南宫28加拿大软件

光华讲坛——社ng28南宫国际app名流与企业家论坛第6973期

主题：Integrative Analysis with Tweedie Model for Multisource Insurance Data under Data Protection Constraints多源异构保护约束保险数据下Tweedie模型的多维度整合建模分析

主讲人：山东财经大学保险南宫28加拿大软件院长助理 张鹏成副教授

主持人：金融南宫28加拿大软件 张晓涵老师

时间：6月26日11:00-12:00

地点：柳林校区格致楼317

主办单位：金融南宫28加拿大软件、中国金融研究院 科研处

主讲人简介：

张鹏成，副教授，硕士生导师，山东财经大学保险南宫28加拿大软件院长助理，山东省泰山学者青年专家。博士毕业于墨尔本大学精算学专业，研究领域涉及保险精算，应用统计，机器学习。近年来在Insurance: Mathematics and Economics、ASTIN Bulletin: The Journal of the IAA、Scandinavian Actuarial Journal、North American Actuarial Journal等国际顶尖精算期刊上发表了多篇学术论文，主持国家自然科学基金青年项目1项、山东省自然科学基金青年项目1项。

内容提要：

This lecture focuses on the prevalent challenges in accurate insurance ratemaking, which arise from data decentralization, stringent privacy supervision, and prominent inherent risk variability across business units. To tackle the above problems, this lecture presents an integrated regression framework with embedded data protection constraints tailored for the Tweedie compound Poisson model, a mainstream model widely adopted for modeling aggregate claim loss data with mixed discrete and continuous distributions. The proposed framework enables multi-source information integration merely through local estimators, Hessian matrices and score vectors without transferring raw underlying data, fully complying with current data privacy and compliance regulations. To achieve model regularization, a difference penalty term is introduced to reduce the coefficient discrepancies of the same covariate across different business units and explore the latent grouping structure of data. Meanwhile, a sparsity penalty term is incorporated to realize efficient and accurate variable selection. A scalable Alternating Direction Method of Multipliers (ADMM) is adopted to efficiently solve the optimization model. Simulation results verify that the proposed integrated regression framework significantly outperforms local Lasso models and achieves performance comparable to centralized modeling in all evaluation metrics. The empirical analysis based on the Medical Expenditure Panel Survey (MEPS) further validates the scientificity and practicality of the proposed framework.

本场悟空体育将围绕当下保险精准费率厘定面临的各类现实难题展开，重点剖析数据分散、隐私监管收紧、各业务单元风险异质性突出等核心行业痛点。为有效解决上述问题，悟空体育将介绍一套适配特威迪复合泊松模型、内嵌数据保护约束的整合回归分析框架。其中，特威迪复合泊松模型是拟合兼具离散与连续混合分布特征的累计理赔损失数据的主流模型。该创新框架仅通过局部估计量、海森矩阵与得分向量即可完成多源数据信息融合，无需传输原始底层数据，完全适配当下数据隐私合规监管要求。为完善模型正则化约束，该框架引入差值惩罚项，有效缩减同一协变量在不同业务单元的系数偏差，精准挖掘数据潜藏的分组结构；同时搭配稀疏惩罚项，完成高效、精准的变量筛选。研究采用可扩展的交替方向乘子法（ADMM），高效求解该优化模型。仿真实验结果证实，该整合回归框架性能显著优于局部套索模型，各项评测指标均与中心化建模水平持平。基于医疗支出面板调查数据库（MEPS）的实证分析，进一步验证了该框架的科学性与实用性。