Penalized regression is steadily becoming a mainstream application in ratemaking. There is a momentum in the insurance space with innovation in research, software and production on penalized techniques that build and innovate the standard GLM models.

Historically, this technique was popularized by the machine learning literature, and how it is taught is not synchronized with how actuaries approach modeling to solve insurance problems. This should not be the case: the penalized framework is versatile and allows to effectively solve many insurance use cases that are currently tackled via established techniques.

First, we will see how, practically and theoretically, Penalized Regressions are effectively Credibility procedures and allow to blend GLM with credibility to reduce overfitting and improve a model’s ability to generalize.

On the other hand, we will display how Penalized regression can be thought of as GBMs, a powerful but yet completely black box modeling technique. 

Since Penalized regressions can effectively tie standard Credibility practices and incorporate some of the benefits of GBMs, this presentation aims to contribute the diffusion of these techniques as solid alternatives to standard GLMs for ratemaking.