Multitask Learning in Tweedie Generalized Linear Models for Insurance Ratemaking

Thomas M. Menino Convention & Exhibition Center 

Insurance ratemaking often requires fitting separate generalized linear models for multiple loss outcomes, such as perils or coverage types, leading to duplicated effort and limited insight into dependence across outcomes. We propose a multitask learning framework for jointly modeling multiple insurance loss responses within a single Tweedie generalized linear model. The approach embeds a regularized multivariate Gaussian regression step within the Iteratively Reweighted Least Squares algorithm (IRLS), allowing dependence across responses to be captured while accommodating semicontinuous loss data. An elastic net penalty is incorporated to address correlated predictors and perform variable selection. Through simulation studies, we demonstrate improved predictive performance and parameter recovery relative to fitting independent models, particularly under multicollinearity. An application to reinsurance data illustrates how the proposed framework enhances interpretability of relationships among loss types while substantially reducing computational time and modeling effort.

ratemaking

multiple perils

Tweedie

elastic net

generalized linear model 

Casualty Actuarial Society