New algorithms and new data for insurance : impact of machine learning techniques in insurance ratemaking

保险公司必须要求确定保费水平的合同的精算估值，是基于旨在正确反映每个投保人风险的“费率制定变量”。例如，年轻司机在汽车保险中的索赔频率(平均)是他们父母的两倍。因此，保险公司长期以来一直将司机的年龄作为评级变量。新数据来源的出现将有助于完善保险公司的概况，从而有助于评估其自身的风险。GLM型模型的经典使用必须进行调整，以纳入如此大量的数据。例如，正在进行的将通过GPS盒获得的驾驶经验数据纳入汽车保险的工作。这项提议的研究领域之一将是开发和实施创新的统计方法，以更好地了解每个被保险人的适当精算风险。除了经典的经济考虑因素外，精算师还必须考虑与保险样本类型相关的统计方面的问题。特别是，在竞争环境中，精算师用来评估合同价值的数据应该反映出选择的偏见。因此，如果保险公司采取高费率政策来吸引某些类型的风险，它们就会流向竞争对手，这些保险公司将只有很少的样本来评估这些合同。超细分将创造非常不同的投资组合，一些类别在一些公司几乎不存在。通常情况下，困难在于估计每个被保险人的道德风险，同时消除因逆向选择而产生的偏见。目前，保费和保险费的计算只是简单地使用前几年观察到的索赔数据，通常使用标准的计量经济模型。在竞争的环境中，应该综合竞争对手的价格作为决定购买保险单的解释变量。我们发现自己处于这样一种情况：计量经济学模型处于竞争之中，并受到竞争的影响。市场上的价格建模必须使用非合作博弈的工具理论和计量经济模型。这项研究计划的目标之一也将集中于将竞争对手的价格整合到差饷制定中。精算估值是基于大数定律和投保人之间的风险分担。公司试图建立风险互助社会，将所有具有共同特征的投保人聚集在一起。正如人们常说的，保险是“多数人对少数人的不幸的贡献”。但是，这种风险分担的概念与保险合同的分割和(高度)个性化原则兼容吗？这项研究提案的目标之一是了解在必要的风险定价方面的影响，以及投保人分割和风险分担之间的平衡。

Actuarial valuation of contracts for setting premium levels that insurance companies must ask, is based on 'ratemaking variables' intended to properly reflect the risk of each insured. For example, the frequency of claims for the young drivers in motor insurance is (on average) twice as high as for their parents. So insurers have long used the age of the driver as a rating variable. The presence of new data sources will help to refine profiling insured and thus the assessment of their own risk. The classic use of GLM-type models must be adapted to incorporate this large amount of data. Examples ongoing work on the inclusion of data related to driving experience, obtained through GPS boxes in motor insurance. One of the research areas of this proposal will be to develop and implement innovative statistical methods to better understand the proper actuarial risk for each insured. In addition to classical economic considerations, actuaries must consider statistical aspects related to the typology of the insured sample available. In particular, in a competitive environment, the data used by actuaries to value the contracts should reflect selection bias. Hence, if an insurer conducts high tariff policy to attract certain types of risk, they will go to competitors, and these insurance companies will then have a very small sample to value these contracts. Hyper-segmentation will create very heterogeneous portfolios, with some categories little present in some companies. Conventionally, the difficulty lies in estimating the moral hazard of each insured while eliminating bias due to adverse selection. Currently the calculation of premiums and insurance premiums is simply using claims data observed in previous years, usually using standard econometric models. In a competitive environment, should be integrated competitors' prices as an explanatory variable in the decision to purchase an insurance policy. We find ourselves in a situation where econometric models are in competition and are influenced one. Price modeling on a market must be done using tools theory of non-cooperative games, and econometric models. One of the goal of this research program will also focus on the integration of competitors price in ratemaking. Actuarial valuation is based on the law of large numbers and the pooling of risks among the insured. Companies try to establish risk mutual societies, grouping all insured with common characteristics. As usually said, insurance is 'the contribution of the many to the misfortune of the few'. But is this concept of risk pooling compatible with the principles of segmentation and (hyper-)individualization of insurance contracts? One of the goal of this research proposal is to understand the impacts in terms of pricing of the necessary the balance between segmentation of policyholders and risk pooling.