报告时间：2023年4月20日

               上午11:00-12:00

报告地点：中国人民大学明德主楼1016

              （腾讯会议ID：748-109-666）

报告嘉宾：杨亮

报告主题：Multi-line Dependent Reserving Based on Generative Neural Network Algorithm

报告摘要

Multi-line Dependent Reserving Based on Generative Neural Network Algorithm

Reserving is the primary means for property and casualty insurance companies to deal with the risk of loss, and it is crucial for risk management and asset allocation. Due to its practical significance and technical difficulties, multi-line reserving has been a focus of attention in the academic community. Among the existing multi-line reserving methods, some model assumptions are too strict, such as the single-line distribution assumption and the assumption of interdependence between multiple lines, which need to be set in advance. Moreover, the model prediction is mostly concentrated on a single central prediction, ignoring the comprehensive distribution prediction of the research object. Therefore, this paper first uses datadriven binning strategy to optimize the smoothing terms of GAM through unsupervised learning, and constructs a GLM_bin model to realize the initial regression of singleline outstanding claim reserves. Secondly, considering the problem of interdependence between multiple lines, the GNN algorithm is used to jointly model the GLM_bin prediction residuals of different lines, which achieves the secondary calibration of the single-line outstanding claim reserving result. Finally, the joint distribution of multiline reserving results is obtained. The results show that compared with the traditional multi-line Copula method, the "GLM_bin + GNN" method not only effectively weakens the model assumptions but also provides powerful support for comprehensive risk management by obtaining joint distribution. Moreover, it improves the accuracy and model robustness of the existing reserving methods, and provides a new perspective for multi-line reserving problems.

keyword:

Non-life insurance; Reserving; Dependent risk measurement；Copula；GNN

作者简介

杨亮，西南财经大学,金融学院副教授。研究方向：非寿险精算费率厘定、准备金评估、风险管理与评估、机器学习算法在车辆网大数据上的应用等。已在《统计研究》《数量经济技术经济研究》《中国软科学》《经济学家》《系统工程理论与实践》《Insurance:Mathematics and Economics》等国内外核心期刊发表论文 10 多篇，主持或参与国家级及省部级项目多项