Credit decisioning with explainability
Gradient boosting and similar models are developed alongside interpretable scorecards; explanations of individual decisions are produced and tested for stability before any use in lending.
GRA · Banking Nexus
空间金融资本与危害、风险暴露和转型的交汇之处
银行业中的机器学习
Artificial intelligence and machine learning now support credit underwriting, fraud and financial crime detection, 早期预警 of borrower stress and supervisory analysis. In banking they are both a tool and a source of risk: their outputs must be explainable, fair, robust and validated before they influence consequential decisions about customers or capital.
01目录
Machine learning in banking in five parts.
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 financial stability, capital markets and investment and climate and hazards, 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.
Early warning and transaction monitoring
Transaction and account data are used to flag borrowers showing early signs of distress and patterns associated with fraud, with every alert reviewed by a person.
Text analysis for risk and supervision
Language models and text classifiers extract risk signals from filings, incident reports and supervisory documents, with outputs checked against human review.
03监测
模型训练。
团队遵循该方法时会使用的界面：其指标、状态与决策。名称为示例，数值仅作示意。
模型训练
Banking · Machine learning in banking · Run R-5640
某一切片上的漂移GRA · Banking Nexus
按切片评估
| 全部 | 0.89 | 0.05 | 5695 | 通过 |
| 区域 A | 0.85 | 0.06 | 1734 | 审核 |
| 区域 B | 0.68 | 0.09 | 636 | 审核 |
| 稀有类别 | 0.71 | 0.11 | 144 | 低于标准 |
方法卡
- 方法
- 人工智能与机器学习
- 拆分
- 基于时间，无信息泄漏
- 功能
- 56
- 标签
- 模型估算
人口稳定性
| 数据 | 角色 | 标签 |
|---|---|---|
| Credit application and performance data | Training and validation | 观测 |
| Transaction records | Monitoring and 早期预警 features | 观测 |
| Public filings and supervisory publications | Text corpora | 已发布统计 |
| Cash-flow and alternative data used under consent and law | Additional borrower features | 已报告 |
控制措施
- 外部资产下降
- 2.0%
- 杠杆率（资产与权益之比）
- 10 倍
- 银行间资产占比
- 10%
- 银行间债权的回收
- 40%
- 体系内的银行
- 20
- 核心银行
- 占银行的 20%
读数
- 初始冲击造成的权益损失
- 18.0%
- 传染后的权益损失
- 45.0%
- 传染新增（DebtRank）
- 27.0 个点
- 违约银行
- 0
数字、方法与来源
| 轮次 | 系统权益损失（%） | 违约银行 |
|---|---|---|
| 0 | 18.00 | 0 |
| 1 | 28.80 | 0 |
| 2 | 35.28 | 0 |
| 3 | 39.17 | 0 |
| 4 | 41.50 | 0 |
| 5 | 42.90 | 0 |
| 6 | 43.74 | 0 |
| 7 | 44.24 | 0 |
| 8 | 44.55 | 0 |
| 9 | 44.73 | 0 |
| 10 | 44.84 | 0 |
| 11 | 44.90 | 0 |
| 12 | 44.94 | 0 |
| 13 | 44.96 | 0 |
| 14 | 44.98 | 0 |
| 15 | 44.99 | 0 |
| 16 | 44.99 | 0 |
| 17 | 45.00 | 0 |
| 18 | 45.00 | 0 |
| 19 | 45.00 | 0 |
| 20 | 45.00 | 0 |
| 21 | 45.00 | 0 |
| 22 | 45.00 | 0 |
| 23 | 45.00 | 0 |
| 24 | 45.00 | 0 |
| 25 | 45.00 | 0 |
| 26 | 45.00 | 0 |
| 27 | 45.00 | 0 |
| 28 | 45.00 | 0 |
| 29 | 45.00 | 0 |
| 30 | 45.00 | 0 |
方法
- DebtRank Battiston 等（2012），Scientific Reports 2:541
- 差异化 DebtRank Bardoscia and others (2015), PLOS ONE 10
- 金融网络中的传染 Gai and Kapadia (2010), Proc. R. Soc. A 466
- 金融网络中的清算 Eisenberg 和 Noe（2001），Management Science 47
- 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
- 单因素敏感性（龙卷风图） Saltelli and others (2008), Global Sensitivity Analysis: The Primer, Wiley
数据
No observed data enters this model: it runs on the teaching values set by the controls, to be replaced with a programme’s own evidence.
贯穿六大支柱，由 Nexus Consortium 牵头
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 hold-out | Models are evaluated on periods later than the training data, never on random splits alone |
| Fairness testing | Outcomes are compared across groups using several fairness definitions, and the trade-offs are reported |
| Explanation stability | Explanations are checked for consistency across similar cases and over time |
| Drift monitoring | Input and output distributions are tracked against thresholds that trigger review and retraining |
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。
07风险暴露、损失与资本
What machine learning in banking 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。
情景路径
How machine learning in banking could be stressed under the scenario families central banks and supervisors use.
| 情景系列 | 转型风险 | 物理风险 | 压力测试问什么 |
|---|---|---|---|
| 有序 | 早期、渐进的政策 | 较低 | 账面能否承受持续的重新定价？ |
| 无序 | 迟缓而突然的政策 | 中低 | 在突然重新定价和信用迁移下，什么会出现问题？ |
| 温室世界 | 有限政策 | 高且上升 | 实体损失集中在抵押品和保险的哪些环节？ |
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
10已关联
本页在整个联合体中通向何处。
11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）Banking Nexus金融与保险准备度情景与压力测试
- 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。
- Zenodo开放存缴存缴并提交到社区用您的 ORCID iD 在 Zenodo 上发布您自己的记录，然后提交到那里的 Reports 社区。由策展人审核；DOI 仍归您所有。
- 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
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。