The insurance regulatory regime introduced in the European Union by the “Solvency II” Directive 2009/138, that has become applicable on 1 January 2016, is aimed to safeguard policyholders and beneficiaries by requiring insurance undertakings to hold own funds able to cover losses, in excess to the expected ones, at the 99.5% confidence level, over a 1‐year period. In order to assess risks and evaluate the regulatory Solvency Capital Requirement, undertakings should compute the probability distribution of the Net Asset Value over a 1‐year period, with a financially inspired market consistent approach. In life insurance, given the peculiarities of the contracts, the valuation of the Net Asset Value distribution requires a nested Monte Carlo simulation, which is extremely time‐consuming. Machine learning techniques are considered a promising candidate to reduce the computational burden of nested simulations. This work investigates the potential of well‐established methods, such as deep learning networks and support vector regressors, when applied to the valuation of the Solvency Capital Requirement of participating life insurance policies, by empirically assessing their effectiveness and by comparing their efficiency and accuracy, also w.r.t. the “traditional” least squares Monte Carlo technique. The work aims also to contribute to the global process of renewal of the European insurance industry, where Solvency II has made the board of directors fully responsible of the choice of evaluation techniques and algorithmic processes, under the periodic monitoring of national supervisory authorities.

中文翻译：

---

嵌套随机模拟中的人寿保险中的机器学习技术

欧盟于2009年1月1日生效的“偿付能力II”指令2009/138引入的保险监管制度旨在通过要求保险企业持有能够弥补损失的自有资金来保护保单持有人和受益人。在1年期间内，以99.5％的置信水平超出了预期的水平。为了评估风险并评估法规的偿付能力资本要求，企业应该使用受财务启发的市场一致方法来计算1年期间净资产价值的概率分布。在人寿保险中，鉴于合同的特殊性，净资产价值分配的估值需要嵌套的蒙特卡洛模拟，这非常耗时。机器学习技术被认为是减少嵌套模拟的计算负担的有前途的候选者。这项工作通过经验评估其有效性并比较其效率和准确性，从而研究了成熟的方法（如深度学习网络和支持向量回归器）在参与人寿保险单的偿付能力资本要求的估值中的潜力，也有“传统的”最小二乘蒙特卡罗技术。这项工作还旨在为欧洲保险业的全球更新进程做出贡献，在国家监管机构的定期监督下，Solvency II已使董事会完全负责评估技术和算法流程的选择。