EcoSta 2022: Start Registration
View Submission - EcoSta2022
A0704
Title: Double boosting of mean and dispersion in Tweedie's compound Poisson model with pre-defined base learners Authors:  Guangyuan Gao - Renmin University of China (China) [presenting]
Abstract: Tweedie's compound Poisson model is a widely used method for predicting insurance loss. It is often necessary to model both mean and dispersion of insurance loss in Tweedie's compound Poisson model under the framework of double generalized linear models. However, the double generalized linear model is restricted to the linearity of covariates, which requires deliberate feature engineering. We propose a double boosting for joint modelling both mean and dispersion. Most boosting algorithms cannot facilitate random effects or spatial effects which often appear in insurance loss prediction. Thus, in the double boosting, we pre-define suitable base learners for different types of covariates. We conduct simulated data analysis and a real data analysis to illustrate the proposed method.