讲座题目：Regression Tree Credibility Model

讲 座 人： Chengguo Weng Department of Statistics and  Actuarial Science University of Waterloo, Canada

讲座时间：2017年6月22日（周四）下午15:00  -16：30

讲座地点：浙江工商大学综合楼601会议室

主讲人简介：

 Chengguo Weng is an Associate Professor of Actuarial Science at the University of Waterloo. He has been working as an Assistant Professor in the School of Statistics and Mathematics at the Zhejiang Gongshang Universtiy during July 2004 - April 2005. He received a Ph.D. in Actuarial Science from the University of Waterloo, a Master of Mathematics and a Bachelor of Science (both in Statistics) from Zhejiang University. His research interests span a wide range of topics in actuarial science and finance, in both theoretical and applied aspects. He has published thirty papers on internationally renowned journals in relevant areas. His latest research focuses on optimal decision, stochastic modelling and predicting problems from the fields of insurance and finance. His research team are currently working on (1) Predictive analytics in insurance and risk management; (2) Portfolio selection in high dimensional settings; (3) Actuarial risk management with basis risk; (4) Portfolio selection based on performance measures; (5) Statistical inference for general stochastic optimization problems.

摘要：

Credibility theory is one of the cornerstones in actuarial  science and has been widely applied for insurance premium prediction. In this  talk I will introduce our research for an SOA-funded project jointly with Dr.  Liqun Diao (University of Waterloo). We bring regression trees techniques into  the credibility theory and propose a novel credibility premium formula, which we  call regression tree credibility (RTC) premium. The proposed RTC method  first recursively binary partitions a collective of individual risks  into exclusive sub-collectives using a regression tree algorithm based on  credibility loss, and then applies the classical Bu¨hlmann-Straub credibility  formula to predict individual net premiums within each sub-collective. The  proposed method effectively predicts individual net premiums by incorporating  covariate information in a very flexible way, and it is particularly  appealing to capture various non-linear covariates effects and/or interaction  effects because no specific regression form needs to be  prespecified in the method. Our proposed RTC method automatically  selects influential covariate variables for premium prediction with  no additional ex ante variable selection procedure required. The superiority in  prediction accuracy of the proposed RTC model is demonstrated by extensive  simulation studies.