Credibility and Machine-Learning: Two Faces of the Same Coin

Generalised Linear Models (GLMs) are very popular among the actuarial community because of their key strengths (simplicity, very strong statistical foundations, and transparency), but they suffer from well known limits, foremost among which is their inability to incorporate Credibility-like assumptions. This limit makes it necessary for the modeler to invest significant efforts in data preparation and variable selection. Machine Learning techniques, on the other hand, provide modelling automation by automatically incorporating regularisation into the models but lack statistical clarity and models transparency, which often make them unsuitable for actuarial modelling.

During his presentation Guillaume will present key concepts and intuitions that demonstrate how regularisation provided by ML algorithms and actuarial credibility are essentially equivalent. By walking through the simple example of Penalised Regression (Ridge and Lasso in particular) he will illustrate how GLM modelling can be improved without losing its key strengths, and how these can be practically leveraged for risk modelling.