Machine learning methods to perform pricing optimization - a comparison with standard GLMs

Speaker(s): Giorgio Alfredo Spedicato (Unipol Group)

Pricing optimization is acquiring increasing importance in personal lines pricing in the most mature insurance market. As the level of competition increases, insurers are forced to optimize their rating and take into account the customer behavior. The dominating methodology is currently GLM based for risk-feature. The authors explored the applicability of machine learning techniques to optimize the proposed premium for new prospective policyholders on a real motor business quotes data set.

Various linear (classical discriminant analysis, elastic net GLM), non-linear (support vector machines, k - nearest neighbours, deep learning), tree-based approaches (C5.0 and random forests) and boosted models (GBM, XGBoost) as well as "Ensembles of Models" have been fitted and compared in term of predictive accuracy and operational requirements. Boosted models and ensembles showed a predictive performance advantage compared to the classical approach, yet their increased complexity induced by their use should be considered.

http://www.variancejournal.org/articlespress/articles/Machine-Spedicato.pdf