Seminar by Dr. Guangyuan Gao from School of Statistics, Renmin University of China

In insurance pricing, models with latent variables are of interest such as Tweedie’s compound Poisson models for pure premium with claim number as latent variable, zero-inflated Poisson (ZIP) models for claim frequency with the source of zero claim as latent variable and finite mixture models for claim severity with component indicator as latent variable. The EM algorithm is a standard model-fitting method for the parametric models with latent variables. However, the EM algorithm is not efficient for fitting additive models since penalized maximum likelihood estimation (PMLE) is used and the smoothing parameters are tuned in each M step. We propose an Expectation-Boosting (EB) algorithm to fit additive tree models with latent variables. The EB algorithm trains only one weak learner at each B iteration in a stagewise fashion. Theory shows the monotone behavior of the likelihood in the EB algorithm. A simulated data example of mixture of Gaussians and a real data example of ZIP model are studied to demonstrate the advantages of the proposed method including variable selection, feature engineering and determination of model structure.

Dr. Guangyuan Gao is an associate professor in School of Statistics at the Renmin University of China. His research area is actuarial science and applied statistics. His work is published in Insurance: Mathematics and Economics, Scandinavian Actuarial Journal, Machine Learning, etc. His research has been supported by the National Natural Science Foundation and the Society of Actuaries Grants.