信贷风控
Design an end-to-end machine learning system for a real-time loan approval/rejection model, such as credit cards. Discuss the infrastructure, features, model, training and evaluation aspects of the system.
1. requirements
functional
stage: 获客、贷前(Loan origination)、贷中(Loan maintenance /servicing)、贷后(Delinquency management/recovery)
types of loans it will support
types of risk it will support -> 决定了是一个什么样的机器学习任务, 影响难点
目标: 识别优劣(风险，需求，价值)
业务指标
non-functional
compliance requirements
scalability goals
reliability, security
2. ML task & pipeline & keys
信贷风控决策流 黑白名单 rule + model binary classification, multi class classification or multi label classification
3. data
user (关系型数据库)
credit 信用
fraud 欺诈
log (分布式文件系统)
label
滚动率(Roll Rate)、vintage
从数据指标中发现风险点
信贷: 逾期不放贷款的就是黑样本
需求类: 履约风险: 履约能力:
外部数据
因合规要求，很多互联网大数据只能从征信平台以评分的形式引入
4. feature
feature engineering is the art
user
ID/Address Proof: Voter ID, Aadhaar, PAN Card
Employment Information, including salary slips
Credit Score
Bank Statements and Previous Loan Statements
反欺诈重要feature，如ip，device_id，idfv，phone number
graph
行为序列
compliance -> buy some user feature
5.model
解释性
Credit Scoring Models 评分卡
LR
GBDT
NN
Probabilistic Calibration
6. evaluation
offline
准确率、AUC、Log Loss、Precision、Recall
Kolmogorov-Smirnov，风控常用指标
7. deploy & serving
feature service
prediction service
8. monitoring & maintenance
Approval Rate
9. QA & optimization
cold start
profit/revenue come from which (credit score) part of customers, risk come from which part of customers
reference
Last updated
