关键词: 地震指数保险, 损失预测, 决策树, 多任务学习, 保险设计和定价

Abstract: The earthquake index insurance (EQII) is an important part of China’s catastrophe insurance system. However, due to the low frequency, high loss, and strong randomness of earthquake disasters, the design of EQII compensation structure often faces problems such as insufficient sample size and inaccurate prediction of losses, which can lead to serious basis risk. This paper proposes a novel method of designing index insurance. We first combine the training results of the generalized additive model and the structure of the decision tree, discretize the continuous index, and simultaneously improve the robustness and practical feasibility of the earthquake loss prediction and the index insurance design. Based on the idea of transferred learning, the multi-task improved decision tree with a soft-parameter sharing structure is proposed, thus realizing the data augmentation of earthquake losses in different regions. Using China earthquake loss data between the years 1949 and 2019, this paper taking Yunnan, Xinjiang, and Tibet-Qinghai-Gansu-Ningxia as the three target areas, verifies the advantages of the novel method in reducing the out-of-sample basis risk of EQII and applies the actuarial “frequency-severity” framework to calculate the premium. Our method will provide more theoretical and practical references for the development of China catastrophe insurance system.

Key words: Earthquake Index Insurance, Loss Prediction, Decision Tree, Multi-task Learning, Insurance Design and Pricing