Skip to content

GRA · Insurance Nexus

空间金融资本与危害、风险暴露和转型的交汇之处

保险中的机器学习

Artificial intelligence and machine learning are used in insurance to read images and documents, detect fraud, triage claims, enrich exposure data and segment risk. On this platform they are treated both as methods and as a source of risk to be governed, tested and explained.

01登记

Machine learning in insurance on the record.

方法记录STEM 方法
  1. 方法人工智能与机器学习
  2. 应用于Insurance Nexus
  3. 规范页面STEM 公共资源
02申请

人工智能与机器学习在此处的作用。

Artificial intelligence and machine learning develop algorithms that learn patterns from data to classify, predict or generate outputs. They are increasingly used for hazard detection, forecasting and decision support, and are themselves a source of risk. In insurance and catastrophe, climate and hazards and health and biosecurity, machine learning can scale detection and forecasting across interconnected 系统. Its own failure modes must be weighed when automated outputs inform consequential decisions. The canonical description is on STEM Commons.

Exposure enrichment from imagery

Models extract roof type, building footprint and vegetation proximity from aerial and satellite imagery to fill gaps in exposure data, with accuracy measured against inspected properties.

Fraud and claims triage

Classifiers score claims for complexity or fraud indicators to route them for review, with false positive rates, appeal outcomes and group disparities monitored beside detection.

Fairness testing of pricing models

Pricing and underwriting models are tested for disparate outcomes and proxy effects across groups, and less discriminatory alternatives are compared for predictive performance.

03损失模型

损失超越、年均损失和可能最大损失。

处于覆盖范围内的投资组合每年预期损失,以及保险公司和贷款机构定价所依据的 1/100 和 1/250 水平下的预期损失,均来自每个 巨灾模型 所用的同样三个模块。教学模型,所用数值为注明的示意值,并非预测。

控制措施

典型年度危害强度(所在地)
1.00 强度指数
极端值的扩散(尺度)
强度指数 0.80
损失一半价值时的强度
4.00 强度指数
脆弱性的离散程度
0.55
危害足迹中的价值占比
40%
防护标准(重现期)
50 年

读数

年均损失
占价值的 0.51%
百年一遇(1/100)的损失
占价值的 24.5%
250 年一遇(1/250)的损失
占价值的 28.4%
百年一遇(1/100)的强度
强度指数 4.68
数字、方法与来源
重现期(年)强度(强度指数)损失(占价值的 %)
10.090.00
21.270.00
41.910.00
72.470.00
133.000.00
243.510.00
444.020.00
824.5323.55
1545.0326.44
2875.5328.87
5366.0330.88
1,0006.5332.53

方法

  1. Gumbel 极值危害 Gumbel (1958), Statistics of Extremes, Columbia University Press
  2. 对数正态易损性(脆弱性) FEMA (2003), HAZUS-MH Technical Manual
  3. 超越概率曲线与年均损失 Grossi and Kunreuther (2005), Catastrophe Modeling, Springer
  4. 堤防效应 Di Baldassarre and others (2013), Socio-hydrology, HESS 17
  5. 不确定性扇形图(参数蒙特卡罗) Spiegelhalter, Pearson and Short (2011), Visualising uncertainty about the future, Science 333
  6. 单因素敏感性(龙卷风图) Saltelli and others (2008), Global Sensitivity Analysis: The Primer, Wiley

数据

The model runs on the teaching values set by the controls. To calibrate it for a place, read it against these admitted layers and series, which this page’s map can draw:

  • 河流 Natural Earth
  • 城市范围 Natural Earth
  • 城市 Natural Earth
  • 过去一周 M4.5+ 地震(USGS) 美国地质调查局 · 登记 usgs.earthquake.summary-feeds
  • 降雨率,最近半小时(IMERG) NASA GPM IMERG Early · 登记 nasa.gibs.context-live-layers
  • 河流流量预报,30 天及范围(GloFAS via Open-Meteo) Open-Meteo.com(CC BY 4.0)、Copernicus GloFAS · 登记 open-meteo.public-pages
  • 建成区与居民点范围,2010 年(HBASE,Landsat) CIESIN/Columbia、NASA SEDAC HBASE · 登记 sedac.context-urban-layers
损失超越概率、年均损失和可能最大损失。示意性教学模型:它依据控件设定的数值运行,而不是实测数据。方法:Gumbel 极值危害(Gumbel(1958),Statistics of Extremes,Columbia University Press);对数正态脆弱性曲线(FEMA(2003),HAZUS-MH Technical Manual);超越概率曲线与年均损失(Grossi 和 Kunreuther(2005),Catastrophe Modeling,Springer)。
04数据

该方法使用的输入。

本平台上人工智能与机器学习所使用的数据集与记录,注明各自的作用及其数值所附的标签。来源列于 来源登记册.

数据角色标签
Aerial and satellite imageryExposure attributes观测
Labelled claims historiesTraining and testing已报告
Property inspection recordsGround truth for imagery models观测
Demographic reference dataFairness testing估算
05风险暴露、损失与资本

What machine learning in insurance means for exposure, loss and capital.

风险官、监管者或财政部会要求的风险暴露、损失、情景和资本视图,以及它们据以报告的框架。结构遵循监管和精算实务。数值为说明性:绝非价格、评级或建议。

各重现期的风险暴露与损失

How losses from machine learning in insurance grow with severity, gross and net of reinsurance, against capital.

1/10(10 年一遇)1/50(50 年一遇)1/100(100 年一遇)1/250(250 年一遇)总额扣除再保险后的净额800
重现期年概率总损失净损失净额,资本 = 100
1/10(10 年一遇)0.10141010
1/50(50 年一遇)0.02332121
1/100(100 年一遇)0.01462727
1/250(250 年一遇)0.004683939
年均损失3.12.22.2
一个示例投资组合的示意性损失指数(资本 = 100):保险公司和监管机构所读的超越概率形式。并非任何机构的数据,也非价格或评级。
NGFS 气候情景 · 央行与监管机构绿色金融网络IFRS S2 气候相关披露 · ISSBTCFD 建议 · 金融稳定理事会 TCFD有效管理气候相关金融风险的原则 · 巴塞尔银行监管委员会
06验证

结果如何核查。

模型在尊重空间、时间与群体结构的留出数据上进行评估。在测试准确性的同时,也测试稳健性、校准与公平性。

核查如何开展
Temporal and held-out validationModels are tested on later periods and unseen regions, not only on random splits
Calibration and drift monitoringPredicted probabilities are compared with outcomes over time, and drift triggers recalibration
Fairness and explainability reviewOutcome disparities and explanations are reviewed before deployment and periodically after
人工监督Decisions adverse to customers have a documented human review and an appeal route

当部署条件与训练数据不同时,性能会下降。模型可能固化偏见,并在罕见事件上毫无明显预警地失效。

08管理责任

由谁维护该方法。

人工智能与自主系统 为 Consortium 管理人工智能与机器学习,并审查其在各平台上的应用情况。

10问题

问题与解答。

在该领域,人工智能与机器学习有哪些局限?
当部署条件与训练数据不同时,性能会下降。模型可能固化偏见,并在罕见事件上毫无明显预警地失效。
结果如何核查?
模型在尊重空间、时间与群体结构的留出数据上进行评估。在测试准确性的同时,也测试稳健性、校准与公平性。
本页面上的数值是真实的吗?
The data names, roles and checks are real descriptions of practice. The canvas shows a design of the method’s use, not a run; no result is reported here.
该方法还应用于何处?
维持社会运转的各个 规范方法页面 列出应用该方法的每个平台。
报告贡献

从本页贡献证据。

报告 is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.

全球风险联盟(Global Risks Alliance)Insurance Nexus金融与保险准备度全球部门风险与韧性
Insurance Nexus

向联盟提出风险暴露或资本问题。

信息和咨询现已开放。每项咨询都会获得一个编号,并告知下一步;申请、付款、算力分配和注册只有在相应团队、条款和审查能力就绪时才会开放。