GRA · 金融体系风险
空间金融资本与危害、风险暴露和转型的交汇之处
AI 与模型风险
Financial institutions rely more and more on models they did not build and cannot fully see: credit scoring, fraud detection, catastrophe models, trading systems and general-purpose AI from a few providers. The Global Risks Alliance (GRA) examines how model failure becomes financial loss: validation gaps, third-party concentration, weak human oversight. It works on shared questions of method, never on any firm’s own models. Part of Financial-system risk.
记录在案的人工智能与模型风险。
模型失效如何转化为金融损失。
| 渠道 | 冲击如何传导至金融 |
|---|---|
| 验证缺口 | A model performs outside the conditions it was tested on, and losses or mispricing follow before anyone notices |
| 第三方集中度 | Many institutions rely on the same model or AI provider, so one fault or outage is correlated across firms |
| Herding | 在相似数据上训练的相似模型,对同一信号作出相同的反应 |
| 监督失灵 | 在没有人能够理解、质疑或叫停的情况下,自动化输出即被付诸行动 |
| Data drift | 气候、人口或市场变化使历史训练数据难以作为可靠依据 |
渠道描述的是机制,并非实测强度或概率。
Model risk rules changed in 2026.
| 工具 | 状态,2026 年 10 月 | 涵盖内容 |
|---|---|---|
| US revised interagency guidance on model risk management (SR 26-2) | Issued April 2026 by the Federal Reserve, OCC and FDIC; supersedes SR 11-7 | Principles-based, risk-tailored model risk management; generative and agentic AI are stated to be outside its scope |
| PRA SS1/23 模型风险管理原则 | 自 2024 年 5 月起对获得内部模型批准的英国银行生效 | 模型识别、治理、开发、验证和风险缓释措施 |
| EU AI Act, as amended by the AI Omnibus | 信用评分高风险义务现自 2027 年 12 月 2 日起适用 | Risk management, data governance, human oversight and documentation for high-risk AI, including creditworthiness assessment |
| NIST 人工智能风险管理框架 | Voluntary, 2023 | Govern, map, measure and manage functions for AI risk |
When many institutions use the same model.
模型风险 就成为稳定性问题。再保险公司之间通用的巨灾模型、贷款机构之间通用的信用评分以及交易台之间通用的市场数据源,都可能同时朝同一方向变动。GRA 在行业层面研究这种共模风险暴露:哪些模型族是共用的,其输入发生变化时会怎样,以及存在何种独立质疑机制。
Method guidance continues at AI and Financial Model Risk; AI oversight governance at the Global Risks Forum (GRF) AI oversight and accountability.
GRA 共同研究的问题。
- 01
复合冲击下的验证
How should models be tested against conditions outside their history, such as compound physical and operational events?
- 02
提供商依赖
How concentrated is the industry’s reliance on a small number of model and AI providers, and what fails together?
- 03
人工监督
What does meaningful human review look like when a decision relies on a model no one in the room built?
- 04
巨灾模型不确定性
对于金融读者,应如何呈现危害模型之间的差异,而不是将其平均掉?
面向行业的方法,而非对某家公司的评审。
- GRA does not validate, approve or certify any institution’s model, and issues no model rating.
- 其成果为共享的方法和证据,在评审和权利许可的范围内发布。
- No member’s confidential model information is exchanged with another.
- Supervisory expectations remain with supervisors; GRA’s cyber and operational work covers the related third-party questions.
联合体内的同一主题。
Where this page’s references lead.
阅读本页时所对照的机构和文件。状态和版本会变化;每个链接都指向发布机构。
问题与解答。
Will GRA review our institution’s models?
Does GRA set standards for AI in finance?
模型风险专家如何做出贡献?
从本页贡献证据。
Reports 是联合体的季刊。您对该主题的了解可以成为下一期中的证据、案例或更正,署名归您,并在属于您本人作品时以 CC BY 4.0 发布。
- 电子邮件门户 GW-GRA从本页发送证据、更正、案例、方法、审查、翻译或概况投稿。它会带上本页标签送达,由负责的编辑接收。
- ZenodoOpen depositDeposit it and submit to the communityPublish your own record on Zenodo with your ORCID iD, then submit it to the Reports community there. Curators review it; the DOI stays yours.
- 报告全球风险联盟(Global Risks Alliance)查看所有贡献方式贡献类型、各自的评审方式、您获得的署名、许可协议,以及绝不接受的内容。
切勿发送个人数据、敏感基础设施细节或您无权分享的任何内容。贡献在使用前会经过审查;发送贡献不赋予署名权或编辑权,赞助也绝不能换取研究结论。
法律文件。
本网站上的地图、数据和图表仅作为背景和一般信息展示。它们不是建议、预测或预警;信息的局限见免责声明和数据说明。