This is a shortened version of a presentation that I did at the Society of Actuaries (SOA) Predictive Analytics Symposium 2019. I cover feature importance, why feature engineering is necessary, why models don’t typically extrapolate well, and the types of features that most models cannot engineer on their own.

Note: Apologies, but I do not seem to have that notebook anymore. I had stored it in gdrive and must have deleted it at some point. I’ve since switch to storing everything in GitHub.

source