Radical Raters IX - by Serban Dragne

The death of the GLM has been greatly exaggerated

Stop being mean about GLMs

An insightful exploration of Generalised Linear Models in motor insurance pricing. Serban highlights GLM's superior performance on tabular data while critically examining the limitations of Auto ML solutions. Key points of the talk include:

• GLMs demonstrate exceptional effectiveness, often outpacing alternative modelling approaches.

• Challenging the assumption that models can function seamlessly without customisation, addressing concerns regarding model robustness and data preparation in the insurance domain.

• A comparative analysis of various modelling techniques, such as basic GLMs, FeedForward Neural Networks, and CatBoost is presented, utilising out-of-sample deviance metrics for evaluation.

• Strategies to enhance GLM performance, including interaction terms and automated feature selection, are thoroughly discussed.

• Summary including practical examples and results that solidify his advocacy for GLMs over more complex methodologies.