Credit risk

Credit risk

Credit risk management is essential to protect lenders, ensure financial stability, make informed lending decisions, and encourage responsible borrowing. It also safeguards investors, informs regulatory oversight, influences pricing, and enhances overall risk management practices.

1. Borrower Credit Risk

  • Definition: The risk that a borrower (individual or business) fails to repay a loan or meet debt obligations (interest or principal).
  • Typical Examples:
    • A mortgage borrower defaulting.
    • A corporate bond issuer missing coupon payments.
  • Measurement Tools:
    • Probability of Default (PD)
    • Loss Given Default (LGD)
    • Exposure at Default (EAD)
      • → Together, these feed into expected credit loss or Basel capital calculations.

Expected Loss Framework (EL = PD × LGD × EAD)

Component
Definition
Example
Analytical Role
PD – Probability of Default
The likelihood that a borrower will default on obligations within a given time horizon (typically 1 year).
If PD = 2%, there’s a 2% chance the borrower defaults next year.
Captures credit quality / likelihood of failure.
LGD – Loss Given Default
The proportion of exposure a bank expects to lose if the borrower defaults, after recoveries.
If loan = $100,000 and recovery = $40,000 → LGD = 60%.
Captures severity of loss after default.
EAD – Exposure at Default
The total value exposed to loss when default occurs. For loans, it’s the outstanding balance; for credit lines, it also includes potential future drawdowns.
If a $120,000 credit line has $100,000 drawn and $10,000 expected additional draw → EAD = $110,000.
Captures scale of exposure.
notion image
1️⃣ PD – Probability of Default
定义
  • 衡量借款人在一定时间内(通常 1 年)发生违约的概率。
  • 通常输出一个 0–1 之间的数值,例如 PD = 0.05 表示一年内有 5% 概率违约。
计算方法 & 模型类型
  • 常用 Logistic Regression(解释性强)、Survival Models(考虑时间因素)、Tree-based Models(捕捉非线性)。
  • 输入变量:
    • 借款人财务指标(Debt/EBITDA、Current Ratio、Profit Margin)
    • 信用记录(违约历史、信用评分变化)
    • 行业与宏观数据(行业波动性、GDP 增长率、利率)
    • 定性因素(管理层质量、公司治理)
应用场景
  • 贷款定价(PD 越高,利差越大)
  • 资本计提(Basel 要求)
  • 投资组合信用风险监控
口语示例
PD estimates the likelihood of a borrower defaulting within a given time frame. We typically use logistic regression for interpretability, with predictors like leverage, liquidity ratios, and macroeconomic indicators. This helps us segment risk tiers, price loans appropriately, and determine regulatory capital under Basel rules.
2️⃣ LGD – Loss Given Default
定义
  • 借款人违约时,银行预期损失占违约时暴露额(EAD)的比例。
  • LGD = 1 − Recovery Rate。
计算方法 & 模型类型
  • Beta Regression:适合 0–1 范围数据,解释性强。
  • Tree-based Models:捕捉非线性关系(担保品、司法管辖区、行业)。
  • 输入变量:
    • 担保品类型与估值(CRE、Machinery、Inventory)
    • 借款合同条款(Seniority、Covenants、Guarantees)
    • 借款人财务健康程度(Leverage、Profitability)
    • 法律与市场因素(司法效率、破产法)
应用场景
  • 贷款定价(LGD 越高,贷款利差应更高)
  • 压力测试(在经济下行时 LGD 上升)
  • 资本充足率测算
口语示例
LGD measures the proportion of exposure we expect to lose if the borrower defaults. We model it using beta regression with inputs like collateral type, loan seniority, and jurisdiction. In downturn scenarios, we apply stress multipliers to reflect lower recovery rates.
3️⃣ EAD – Exposure at Default
定义
  • 借款人违约时,银行的总风险暴露额。
  • 对于 Term Loan,通常是剩余本金 + 应计利息;对于 Revolving Facility,要考虑潜在的提款行为。
计算方法 & 模型类型
  • Term Loans:直接计算当前余额 + 应计利息。
  • Revolving Credit Lines:估计 Credit Conversion Factor (CCF),预测违约时可能额外提款比例。
  • 输入变量:
    • 信用额度与已用额度
    • 借款人信用评级
    • 行业与经济环境
    • 历史提款模式
EAD is the total amount we’re exposed to if a borrower defaults. For term loans, that’s straightforward — principal plus accrued interest. For revolving facilities, we estimate a credit conversion factor using historical utilization at default, segmented by borrower rating and sector.

Expected Loss and Applications

#
Application
How Expected Loss (EL) Is Used
1️⃣ Loan Pricing
Banks incorporate EL into loan pricing models to ensure interest rates adequately compensate for the credit risk taken.
Ensures risk-adjusted pricing → borrowers with higher PD or LGD pay higher spreads.
2️⃣ Capital Allocation
Accurate EL estimation helps determine how much capital needs to be reserved to absorb future credit losses.
Supports Basel capital adequacy and internal economic capital allocation (ICAAP).
3️⃣ Performance Measurement
Comparing actual vs. expected losses allows banks to assess model accuracy and portfolio performance.
Enables backtesting and model validation under SR 11-7 governance.
4️⃣ Portfolio Management
Segment-level EL helps identify concentrations, risk hotspots, and underperforming asset classes.
Guides portfolio optimization, risk-based limits, and stress testing.

Managing Credit Risk – Key Techniques

Method
Description
Risk Control Purpose
Collateral
The borrower pledges assets (e.g., real estate, securities) that can be seized in case of default.
Reduces Loss Given Default (LGD) by improving recovery value.
Covenants
Contractual terms requiring borrowers to meet certain conditions (e.g., maintain leverage ratio).
Provides early warning signals and control triggers before default.
Diversification
Spreading exposures across different borrowers, sectors, and regions.
Mitigates concentration risk and portfolio volatility.
Client Selection
Evaluating creditworthiness through financial ratios, credit scores, and qualitative assessments.
Lowers Probability of Default (PD) by screening risky clients.
Credit Limits
Setting exposure caps per borrower, sector, or product.
Prevents excessive exposure to any single counterparty.
Credit Derivatives
Instruments like Credit Default Swaps (CDS) transfer credit risk to another party.
Offloads risk and provides hedging flexibility.
Monitoring & Stress Tests
Ongoing review of portfolio performance and scenario simulations (e.g., recession, rate shocks).
Identifies emerging vulnerabilities under extreme conditions.
Securitization
Pooling loans and selling them as securities (ABS, MBS, CDOs).
Frees up capital and redistributes credit risk to investors.
 
 

2. Counterparty Credit Risk (CCR)

  • Definition: The risk that the counterparty in a financial transaction (like a derivative or repo) fails to meet contractual obligations.
  • Typical Examples:
    • A swap counterparty defaulting before settlement.
    • A securities lending borrower failing to return collateral.
  • Measurement Tools:
    • Current Exposure (CE) and Potential Future Exposure (PFE)
    • Credit Valuation Adjustment (CVA) for pricing CCR into derivatives.

3.Credit Analysis(信用分析)

回答框架:5C’s of Credit
  1. Character(品格):借款人的还款意愿和信用历史如何?
  1. Capacity(能力):借款人是否有足够的现金流来偿还债务?
      • 关键指标:Debt-to-Income、Debt Service Coverage Ratio
  1. Capital(资本):借款人有多少自有资金投入?财务状况是否健康?
  1. Collateral(抵押品):如果违约,有什么资产可以用来抵押?
  1. Conditions(条件):宏观经济环境、行业趋势等外部条件如何?
 

4.Credit Risk Modeling

核心概念:
  • PD (Probability of Default):违约概率,借款人在未来一段时间内发生违约的可能性。
  • LGD (Loss Given Default):违约损失率,一旦违约预计会损失多少钱。
  • EAD (Exposure at Default):违约风险敞口,违约时银行对借款人的风险敞口。
  • Expected Loss (EL) = PD × LGD × EAD
回答框架:
  1. 模型选择:为什么选择某个特定模型(如 Logistic Regression、Gradient Boosting)?
  1. 特征工程:你会选择哪些变量来预测违约?如何处理缺失值和异常值?
  1. 模型验证:如何评估模型的预测能力?
      • 指标:AUC-ROC、Gini、KS
  1. 模型监控:模型上线后,如何监控其表现是否衰退?
“Under Basel’s Internal Ratings-Based (IRB) approach, Expected Loss is defined as:
EL = PD × LGD × EAD.
These three components are modeled separately but calibrated to ensure consistency and conservatism for capital adequacy.”
To estimate long-term, risk-sensitive parameters that feed into Regulatory Capital (RWA), Stress Testing, and Credit Portfolio Management.
 
Credit Risk Modeling
Credit Risk Modeling
一份完全模拟真实商业银行 CRE PD 模型开发文档(MDD)
一份完全模拟真实商业银行 CRE PD 模型开发文档(MDD)
 

5.Regulatory & Compliance(监管与合规)

核心概念:
  • Basel Accords (I, II, III):全球银行监管的核心框架,定义资本充足率与风险权重。
  • CECL (Current Expected Credit Loss):要求银行更早计提预期信用损失准备。
  • IFRS 9
  • Dodd–Frank Act:2008年金融危机后美国推出的重要金融监管法案。
  • SR 11-7 (Fed Guidance)
    • 模型风险管理的黄金标准

1.Basel Accords (I, II, III) — 巴塞尔协议

Definition:
The Basel Accords are international banking regulations developed by the Basel Committee to ensure financial stability by setting minimum capital and risk management standards for banks.
Core Ideas:
  • Basel I (1988): Introduced the concept of risk-weighted assets (RWA) and minimum capital adequacy ratio (8%).
  • Basel II (2004): Expanded into three pillars:
      1. Minimum capital requirements (for credit, market, and operational risk)
      1. Supervisory review
      1. Market discipline
  • Basel III (2010–): Strengthened capital and liquidity standards (e.g., CET1 ratio, LCR, NSFR) after the 2008 crisis.
Key Impact on Modeling:
Banks must develop IRB (Internal Ratings-Based) models — PD, LGD, and EAD — to estimate regulatory capital more accurately.

2.CECL (Current Expected Credit Loss) — 预期信用损失模型

Definition:
CECL is a U.S. accounting standard (ASC 326) introduced by FASB. It requires banks to estimate lifetime expected credit losses on financial assets at origination, rather than waiting for losses to be “probable.”
Key Change:
  • Old rule (Incurred Loss) → recognize loss after a credit event.
  • CECL → recognize loss before a credit event, using forward-looking forecasts.
Modeling Impact:
  • Uses macroeconomic scenarios, probability of default (PD), loss given default (LGD), and exposure at default (EAD).
  • Requires stress testing and scenario sensitivity under different macroeconomic conditions.

4.SR 11-7 (Federal Reserve Guidance on Model Risk Management) — 模型风险管理黄金标准

Definition:
SR 11-7 is the U.S. Federal Reserve’s supervisory guidance that defines what constitutes model risk and how banks must manage it.
Key Principles:
  1. Sound Model Development and Implementation
      • Models must be conceptually sound, well-documented, and tested before use.
  1. Independent Model Validation
      • Models must be validated by independent teams separate from developers.
  1. Governance and Oversight
      • Senior management must ensure an effective model inventory, issue tracking, and periodic review.
Why It Matters:
SR 11-7 is now the foundation of all MRM frameworks in U.S. banks.
Every model—PD, LGD, EAD, or AI—must comply with its lifecycle:
Development → Validation → Approval → Monitoring → Retirement.

关键区别详解

框架
谁用?
目的
核心逻辑
CECL(FASB ASC 326)
美国 GAAP 企业(如招行纽约分行)
财务会计准则:计提贷款损失准备金
一生预期损失(Lifetime ECL),不管风险是否恶化
IFRS 9
国际 IFRS 企业(如欧洲/亚洲银行)
国际财务报告准则
三阶段模型:Stage 1(12个月ECL)、Stage 2/3(一生ECL)
Basel / IRB(巴塞尔协议)
全球银行(监管资本要求)
监管资本充足率
12个月 PD × LGD × EAD(仅用于资本,不用于会计拨备)
1. CECL vs IFRS 9
  • 相同点:都要求“预期信用损失”(不是已发生损失)。
  • 不同点
    • CECL所有贷款从第一天起就计提一生ECL(哪怕客户信用很好)。
    • IFRS 9:分三阶段:
      • Stage 1:信用未恶化 → 计提 12个月ECL
      • Stage 2/3:信用显著恶化或已违约 → 计提 一生ECL
    • ✅ 所以 IFRS 9 更“渐进”,CECL 更“保守”。
💡 你在招行纽约分行做的 CECL 模型,就是按 一生ECL 来算拨备的,不管客户当前是否违约。
2. CECL/IFRS 9 vs Basel
  • Basel 不是用来做会计拨备的!
    • Basel 的 PD/LGD/EAD 是用来计算 最低监管资本(Capital Requirement)。
    • 它只要求 12个月违约概率(即使贷款期限10年)。
    • 而 CECL/IFRS 9 是 会计准则,直接影响利润表和资产负债表(拨备 = 费用)。
  • 数据可以复用,但目的不同
    • 你在开发 CRE 的 PD 模型时,如果用于 Basel 资本,只需12个月PD;
    • 如果用于 CECL,则需建模 一生违约路径 + 宏观情景(如 Moody’s 宏观变量)。
举个例子(CRE 贷款,5年期)
框架
怎么算损失?
CECL
预测未来5年每年的违约概率 → 折现求和 → 一生ECL
IFRS 9
如果客户信用稳定 → 只算第1年ECL;如果信用恶化 → 算5年ECL
Basel
只看未来12个月会不会违约 → 用于计算资本,不管后面4年

6.CECL for CRE Loan Provision

1.CECL 基本原理 (Conceptual Foundation)

核心问题
你需要掌握的要点
面试关键词
What is CECL?
CECL = Current Expected Credit Loss. 自 2020 年起,美国银行必须在资产发放时确认“整个生命周期的预期损失 (lifetime expected loss)”。
“forward-looking”, “lifetime loss”, “expected credit loss”, “FASB ASC 326”
核心逻辑公式
[ ECL = Σ (PDₜ × LGDₜ × EADₜ) / (1+r)ᵗ ]
“PD, LGD, EAD”, “discounted lifetime loss”
与 IFRS 9 区别
IFRS9 = 阶段模型 (12-month vs lifetime),CECL = 一律 lifetime loss。
“Stage-based vs lifetime approach”
监管意义
CECL 属于 US GAAP 会计标准,而不是监管资本要求;但银行内部通常将 CECL 输出与 Basel capital 一并分析。
“Accounting reserve vs capital requirement”

2.CECL 模型核心组件 (Model Components)

组件
含义
驱动变量
你要能解释的内容
PD (Probability of Default)
借款人在未来期间内违约的概率
Risk grade, DSCR, macro factors (unemployment)
如何受经济情景和信用迁移影响
LGD (Loss Given Default)
违约时无法收回的比例
Collateral value, LTV, recovery rate
抵押品价值下降 → LGD 上升
EAD (Exposure at Default)
违约时的敞口金额
Loan balance, credit utilization
EAD 增长 → 拨备上升
Discount Rate (r)
将未来损失折现为现值
Effective interest rate
折现率越高 → 拨备略低

3.数据与输入层 (Data Inputs & Segmentation)

1️⃣ Loan-Level Data(贷款层级数据)
这类数据是模型计算的“底层颗粒度(granularity)”,
主要决定了PD(违约概率)LGD(损失率)、和 **EAD(敞口)**的输入。
字段
中文解释
作用(在CECL模型中)
对拨备的影响
Loan ID
唯一识别每笔贷款
用于追踪单笔贷款的历史表现
非直接影响,但用于识别数据稳定性
Loan Balance (Outstanding Balance)
当前贷款余额
直接决定 Exposure at Default (EAD)
Balance ↑ → EAD ↑ → 拨备 ↑
Loan Type / Product
贷款类别(CRE, Residential, Credit Card, Auto)
区分不同产品的建模参数
不同产品PD/LGD模型不同
Property Type (for CRE)
办公楼、零售、酒店、工业、住宅
不同物业风险特征不同
Office/Hotel 风险高 → LGD ↑
Collateral Value
抵押物当前估值
用于计算 LTV
抵押物下降 → LTV ↑ → LGD ↑
LTV (Loan-to-Value)
贷款余额 / 抵押物价值
衡量借款人杠杆
LTV ↑ → LGD ↑
DSCR (Debt Service Coverage Ratio)
现金流覆盖率 = NOI / 债务服务额
衡量借款人偿债能力
DSCR ↓ → PD ↑
Interest Rate Type
固定/浮动
影响未来现金流折现与风险
利率上升 → 还款压力 ↑ → PD ↑
Maturity Date / Remaining Term
到期日 / 剩余期限
决定 lifetime horizon
期限长 → Lifetime Loss ↑
Origination Vintage (Year)
发放年份
用于分 cohort 分析
新贷款风险更高(无还款历史)
Risk Grade / Rating
内部评级
映射到 PD 模型区间
Rating 下降 → PD ↑
Geography / Region
地理位置
与区域经济数据绑定
弱经济区域 → PD ↑、LGD ↑
2️⃣ Macroeconomic Inputs(宏观经济变量)
CECL 模型的最大特点是“forward-looking”,
所以宏观变量直接驱动 PD 和 LGD 的未来路径(forecast path)
变量
中文解释
模型作用
对拨备的影响
Unemployment Rate
失业率
PD 模型核心驱动(消费与违约强相关)
↑ → PD ↑ → 拨备 ↑
GDP Growth
国内生产总值增长率
反映经济扩张或衰退
↓ → PD ↑(尤其对企业贷款)
CPI / Inflation
通胀率
影响家庭可支配收入
↑ → Payment rate ↓ → PD ↑
Interest Rates (Fed Funds, 10Y)
利率
决定债务负担和折现率
↑ → PD ↑;同时折现率↑ → 拨备略↓(offset)
CRE Price Index / Housing Price Index
房地产价格指数
LGD 预测关键变量
↓ → LGD ↑ → 拨备 ↑
Consumer Confidence Index
消费者信心
信用卡行为的领先指标
↓ → Payment rate ↓ → PD ↑
Unemployment Forecast Horizon
未来失业路径
决定 lifetime PD 曲线
路径越弱 → Lifetime loss ↑
💡 面试说法:
“We feed forward-looking macro variables like unemployment and property prices into PD and LGD models.
If macro assumptions worsen, lifetime expected losses and CECL provisions increase even if portfolio quality stays stable.”

监管与治理要点

  • CECL 模型属于 高风险模型,需纳入 模型风险管理体系(MRM)
  • 需建立清晰的 模型分类决策树(如你设计的工具)。
  • 验证需覆盖:概念合理性、数据质量、实现准确性、持续监控
 

7.Credit card Credit Risk Management

notion image

New account 新户批核阶段

Objective

在发卡时准确评估客户信用风险,平衡获客与违约风险。

Core Logic

  • 决定是否批准、授信额度、初始定价(APR)。
  • 风控重点是 “谁能批 + 批多少 + 批完多久会坏”。

Key Input Data

类别
示例
Applicant Info
Age, income, employment, housing status
Credit Bureau
FICO score, inquiries, tradeline history
Internal Data
Existing relationship, deposit or loan history
Derived Features
DTI (Debt-to-Income), utilization ratio, payment-to-income

Key Metrics

指标
含义
Approval Rate
批核比例
Bad Rate / Early Default Rate
批卡后早期违约比例(如6个月内30+ DPD)
Booked Volume
发卡量 / 新增余额
Average Limit
平均授信额度
Vintage Curve
每期新户的表现随时间变化趋势

2LoD Oversight Focus

风控环节
二线关注点
Approval Policy
是否符合风险偏好(Risk Appetite)
Cutoff & Tiering
是否有文档支持,cutoff合理性
Model Usage
PD模型是否在可用范围内、稳定性是否验证
Exception Tracking
手动批核(override)比例与理由
Vintage Performance
新户分期表现是否符合预测

Customer Management存量账户管理阶段

Objective

在存量阶段平衡风险与收益——控制坏账同时维持健康增长。

Core Logic

  • 定期评估客户行为变化;
  • 动态调整额度、限额、催收与留存策略;
  • 风险点在“老客户变坏 / 信用风险迁移”。

Key Input Data

类别
示例
Behavioral Data
Payment amount, payment rate, utilization, cash advance usage
Account Info
Credit limit, balance, product type
Delinquency History
DPD, roll rates
Macroeconomic Variables
Unemployment, CPI, interest rates
Derived Features
Revolver vs Transactor, balance growth, partial pay ratio

Key Metrics (KRIs)

指标
含义
Delinquency Rate (30+/60+/90+)
存量客户的逾期率
Roll Rate
向更严重逾期滚动的比例
Charge-off Rate
实际核销比例
Utilization Rate
使用额度比,反映压力
Payment Rate / Partial Pay %
还款能力信号
Balance Growth / Limit Usage
组合敞口趋势

2LoD Oversight Focus

主题
二线问题 / 审查角度
Limit Increase Policy
是否对高风险客户过度提额;有无监控提额后 delinquency 变化
Collections Effectiveness
Recovery 是否下降;催收是否公平合规
Early Warning
是否有有效的 partial pay / utilization trigger
KRI Thresholds
逾期率是否超限;阈值是否合理
Provision / CECL Alignment
拨备是否反映组合真实风险
Portfolio Concentration
某产品或地区是否集中度过高

信用卡各种指标

阶段
指标
英文名称
通俗解释
风控意义
1️⃣ 正常使用期
Utilization Rate
Balance / Credit Limit
用掉的额度占比,比如额度1万、用了8千 → 利用率80%
利用率太高 = 压力大、还款风险上升
Payment Rate
Payment / Statement Balance
客户每月还了多少比例,比如账单1000、还800 → 还款率80%
还款率下降是风险前兆
Transactor vs Revolver
Transactor:全额还款;Revolver:分期滚动
Revolver群体收益高但风险高
2️⃣ 逾期初期
DPD (Days Past Due)
从最早未还账单开始算的天数(30 DPD = 拖欠30天)
核心概念,决定逾期阶段
Delinquency Rate (30+/60+/90+)
30+ = 拖30天以上余额占比;60+ = 拖60天以上余额占比
表示不同阶段的风险严重程度
3️⃣ 逾期演变期
Roll Rate
% moving 30→60→90 DPD
上个月30DPD客户中,有多少“滚动”到60DPD
反映风险恶化速度
4️⃣ 坏账阶段
Charge-off Rate
Charged-off Balance / Total Balance
银行认为无法收回的部分(比如逾期120天后)
实际损失指标
Recovery Rate
Amount Recovered / Charged-off Amount
被催回来的坏账比例
衡量催收效果
5️⃣ 拨备阶段(会计层面)
Provision / ECL
Expected Credit Loss
会计上提前为可能坏账计提损失
金融账面健康度

Delinquency Rate逾期

术语
全称
含义(通俗解释)
DPD
Days Past Due
欠款天数 —— 指客户最早未还款的天数。例如:账单到期后拖了30天没还 → 30 DPD。
Delinquent Account
逾期账户
DPD > 0 的账户。
Delinquency Bucket
逾期桶(阶段)
银行常把账户按逾期天数分桶:0(Current)、1–29(Early)、30–59、60–89、90+。
Delinquency Rate
逾期率
某时点逾期账户余额 ÷ 总余额。反映组合中“有问题”的比例。
逾期率的几种版本(常见面试混淆点)
指标名称
英文
计算逻辑
解释
30+ Delinquency Rate
% of balance ≥30 DPD
逾期超过30天的账户余额 / 总余额
早期风险信号
60+ Delinquency Rate
% of balance ≥60 DPD
逾期超过60天的账户余额 / 总余额
严重逾期
90+ Delinquency Rate
% of balance ≥90 DPD
逾期超过90天的账户余额 / 总余额
进入“核销”前阶段
🧠 记法:
DPD 就像“发烧温度”。
30 DPD → 轻微发烧;60 DPD → 症状明显;90 DPD → 基本确诊(要核销)。
SQL / 数学公式举例
SELECT SUM(CASE WHEN dpd >= 30 THEN balance ELSE 0 END) / SUM(balance) AS delinquency_30p, SUM(CASE WHEN dpd >= 60 THEN balance ELSE 0 END) / SUM(balance) AS delinquency_60p, SUM(CASE WHEN dpd >= 90 THEN balance ELSE 0 END) / SUM(balance) AS delinquency_90p FROM loan_portfolio WHERE calendar_month = '2025-10';
Delinquency → Write-off(核销)
通常银行规定:
  • 信用卡账户 90 DPD 或 120 DPD 后视为无法收回。
  • 核销(Charge-off)意味着会计上确认损失,但仍可能继续催收。
Write-off Rate = 核销金额 / 平均余额

Roll Rate(迁移率 / 滚动率)

1.Roll Rate 是什么?为什么重要
“Roll rate measures how quickly accounts move from one delinquency stage to the next — for example, from 30 days past due to 60 or more.
I calculate it as the percentage of accounts that were 30 DPD last month and are 60+ DPD this month.
Rising roll rates mean risk is accelerating, even if total delinquency hasn’t changed yet.
It’s a key early-warning metric we monitor in portfolio risk dashboards.”
Roll rate = 在某个逾期桶(bucket)中的账户,
下一期滚动到更严重逾期阶段的比例。
它回答的问题是:
“上个月逾期30天的客户,有多少在这个月变成逾期60天或更糟?”
这就是风险“恶化速度”的指标。
原因
说明
早期预警
比 delinquency 更早反映风险加速。
预测坏账
未来 charge-off 数量 ≈ 当前余额 × Roll Rate 链式相乘。
催收策略依据
识别最可能继续恶化的客户群。
CECL 模型输入
常用来估算 PD(违约概率)。
2. 一个最简单的例子
假设:

2024-12 月(上个月)

1000 个账户在 30 天逾期(30 bucket)

2025-01 月(这个月)

这 1000 人的去向:
本月状态
人数
Current
200
30
500
60
200
90+
100
那 Roll Rate 就是:
  • 30 → Current = 200 / 1000 = 20%(变好了)
  • 30 → 30 = 500 / 1000 = 50%(没变)
  • 30 → 60 = 200 / 1000 = 20%(恶化)
  • 30 → 90+ = 100 / 1000 = 10%(严重恶化)
👉 最重要的恶化指标是:
30 → 60 = 20%
30 → 90+ = 10%
就是这 1000 个里,
有多少人变得更糟糕了
 
 
3.通用公式
notion image
 
4.SQL 实现逻辑
假设你的表叫 account_monthly,有字段:
  • account_id
  • calendar_month
  • dpd(逾期天数)
SQL 示例:
WITH base AS ( SELECT account_id, calendar_month, CASE WHEN dpd >= 90 THEN '90+' WHEN dpd >= 60 THEN '60' WHEN dpd >= 30 THEN '30' ELSE 'Current' END AS bucket FROM account_monthly ), lagged AS ( SELECT account_id, calendar_month, bucket, LAG(bucket) OVER (PARTITION BY account_id ORDER BY calendar_month) AS prev_bucket FROM base ) SELECT prev_bucket, bucket, COUNT(*) AS num_accounts, COUNT(*) * 1.0 / SUM(COUNT(*)) OVER (PARTITION BY prev_bucket) AS roll_rate FROM lagged GROUP BY prev_bucket, bucket ORDER BY prev_bucket, bucket;

Step 1: Base 表 — 分类逾期阶段

做的事:
把每个账户每月的 dpd(逾期天数)
分成四个阶段(Bucket):
dpd区间
bucket标签
dpd < 30
Current
30 ≤ dpd < 60
30
60 ≤ dpd < 90
60
dpd ≥ 90
90+
👉 这样后面就可以统计这些“状态”之间的变化。

Step 2: lagged 表 — 获取上个月的 bucket

做的事:
LAG() 函数,取出同一个账户上个月的 bucket 状态
account_id
month
bucket(当前月)
prev_bucket(上月)
A
2024-06
30
Current
A
2024-07
60
30
B
2024-07
90+
60
👉 这一步就是在找每个客户的逾期迁移轨迹

Step 3: 计算 Roll Rate(迁移比例)

做的事:
  1. 统计每个逾期阶段的客户从上月 → 本月的迁移数量。
  1. 用分母(同一个上月 bucket 的总数)做归一化,算出比例。

结果长什么样?

prev_bucket
bucket
num_accounts
roll_rate
Current
Current
9200
0.92
Current
30
600
0.06
Current
60
150
0.015
30
60
400
0.40
30
90+
200
0.20
60
90+
300
0.60
5.解读输出
prev_bucket
bucket
roll_rate
Current
30
1.5%
30
60
20%
30
90+
5%
60
90+
50%
这张表告诉你:
  • 有 20% 的30DPD客户在下个月变成了60DPD;
  • 有 5% 直接从30DPD滚到90+;
  • 有 50% 的60DPD客户又恶化到了90+。

Write-off(或 Charge-off)

“Write-off represents the final stage of credit deterioration — when balances are deemed uncollectible and removed from the books.
In credit cards, charge-offs typically occur after 120 or 180 days past due.
I calculate the write-off rate as charge-off amount divided by average balance, and monitor it alongside delinquency and roll rates to see how quickly accounts transition to loss.
As part of second-line oversight, I ensure write-off trends remain consistent with CECL provisions and that recovery performance is tracked to assess collections effectiveness.”
1.概念:Write-off 是什么?
Write-off(或 Charge-off)= 银行在会计上把某笔贷款认定为“无法收回”,
从账面资产中核销掉,但催收仍可能继续。
换句话说:钱可能还在催,但财务上已经认定是“损失”。
监管要求银行不要高估资产价值。
在信用卡领域(无抵押贷款),通常规定:
逾期180天(或120天)未还 → 必须核销。
所以 Write-off 是有制度性触发的,不是随机决定的。
2.公式:Write-off Rate 怎么算?
notion image
举例:
  • 当月平均贷款余额 = $100M
  • 当月核销金额 = $2M
    • → Write-off Rate = 2%
3.数据示例(信用卡组合)SQL 示例
月份
Balance
Charge-off Amount
Write-off Rate
Jan
100M
1.2M
1.2%
Feb
102M
1.5M
1.47%
Mar
104M
2.0M
1.9%
趋势上升 ⇒ 组合坏账加快。
SELECT calendar_month, SUM(CASE WHEN charge_off_flag = 1 THEN charge_off_amount ELSE 0 END) * 1.0 / AVG(balance) AS write_off_rate FROM loan_portfolio GROUP BY calendar_month ORDER BY calendar_month;
4.Write-off 与前面指标的关系
指标
含义
与 Write-off 的关系
Delinquency Rate
当前逾期的比例
是坏账的“存量”;部分会滚到核销
Roll Rate
从轻度逾期到重度逾期的迁移率
决定坏账形成的速度
Write-off Rate
本期真正损失的比例
是坏账的“流量”结果
Recovery Rate
催收追回比例
抵消部分损失,提升净收益
一句话逻辑链:
Utilization↑ → Payment↓ → Delinquency↑ → Roll rate↑ → Write-off↑.
5.Recovery(回收率)
Write-off 之后银行仍可催收。
Recovery Rate = 回收金额 / 核销金额
例如:核销了 $10M,后来催回 $2M → Recovery = 20%。
6.二线 (Second Line) 怎么看 Write-off?
二线关注点
说明
Write-off Trends
是否上升超出风险偏好限额
Roll-to-Write-off Link
是否与逾期趋势一致(数据对齐)
Collections Effectiveness
Recovery Rate 是否下降
Provision Alignment
拨备(CECL)是否充分覆盖未来核销