银行消费风险解决方案部（CRS）损失预测.ppt

Consumer Risk Solution(CRS)Loss Forecast消费风险解决方案部（CRS）损失预测 CRS Decision Analytics SupportCRS 决策分析支持,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,2,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,Agenda 提纲,Loss Forecasting Function 损失预测职能Loss Forecasting Process 损失预测流程Key Factors to Consider in the Consensus Forecast 共识预测中要考虑的关键因素Loss Forecasting Models and Methodologies 损失预测的模型及方法Roll Rate Model 滚动比例模型Delinquency Migration Matrix(DMM)Model 还款逾期移动矩阵（DMM）模型 Capital Allocation Model 资本分配模型Econometric Model 计量经济学模型Account Level Models 账户级模型Other Models/Tools 其他模型/工具Advantages and Limitations of Different Models 不同模型的优点和局限性Model Development and Validation 模型开发和验证Scenario Analysis of the Home Equity Line of Credit(HELOC)Portfolio 住房净值贷款（HELOC）的情境分析Scenario Analysis Methodology 情境分析方法Scenario Assumptions And Simulations on Housing Bubble And Oil Price 住房泡沫和油价的情境假设和模拟Benchmark Reference of Scenario Analysis 情境分析的参考基准,3,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,Provide loss forecasts and asset quality forecasts for the strategic plan and budget plan.提供损失预测和资产质量预测，为制定战略计划和预算计划提供依据。Deliver monthly loss forecast update and asset quality reporting to support enterprise risk department on loss provisions and reserve requirements.提供月度损失预测更新及资产质量报告，为集团风险部门进行贷款拨付和满足准备金要求提供支持。Track and monitor asset quality and portfolio performance against budget plan and Hoshin plan.对照预算计划和 Hoshin 计划，追踪和监督资产质量和资产组合绩效。Recommend risk mitigation strategies through the consensus forecast process.通过共识预测流程提出风险降低战略的建议。Develop statistical models to support loss forecasts and scenario analysis.开发统计模型，为损失预测和情境分析提供支持。Conduct scenario analysis to project potential losses.进行情境分析，以预测潜在的损失。,Consumer Risk Solution Loss Forecasting Functions消费风险解决方案部的损失预测职能,4,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,Consumer Risk Solution High Level Loss Forecasting Process消费风险解决方案部高级别损失预测流程,Run Forecasting Models运行预测模型,Trigger:Monthly Forecast Due触发因素：编制月度预测,Various Model Outputs各种模型输出,Data Warehouse and Systems数据仓库及系统,Corporate ResearchMacro Economic TrendFinancial Planning and Forecast集团研究部宏观经济趋势,财务规划及预测,Generate/UpdateForecast生成/更新预测,Preview Forecastwith CROG与CROG共同预读预测结果,MajorVariance?是否有重大差异？,Yes 是,Research DiscrepanciesUpdated Models研究差异点更新模型,Review Model Outputs And Forecastswith LOB Risk Partners,Finance,CROG,andBusiness Owners与业务线风险伙伴部门、财务部、CROG和业务主管部门共同审查模型的输出和预测,Generate Consensus ForecastUpdateLoss Forecast Database System生成共识预测更新损失预测数据库系统,Produce Forecast Reportsand Asset QualityReports生成预测报告和资产质量报告,Reports报告,End 结束,No 否,Key Factors to Consider in the consensus forecast共识预测中要考虑的关键因素,NOTE:CROG stands for“Credit Risk Operations Group 注：CROG代表”信用风险业务组”,5,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,Model outputs provide a base and should be considered as a input in the forecast process.模型输出提供了一个基础，应作为预测流程的一项输入加以考虑。Business strategies and initiatives and their impacts on the portfolios losses.业务战略和行动计划，及其对资产组合损失的影响。Credit risk operation activities and risk mitigation strategies and their impacts on the actual losses and recovery.信用风险业务活动和风险降低战略，及其对实际损失和回收率的影响。Change in credit risk policies and new product launch and their impacts on the portfolios losses.E.g.,marketing interest-only and Option Adjustable Rate Mortgages(ARMs)信用风险政策的变化和新产品推出，及其对资产组合损失的影响。例如，推销只付利息住房贷款和本息还款可选择的可调整利率住房贷款（ARM）。Macro economic environment and consumer credit trend.宏观经济环境和消费信贷趋势。Impacts from external factors such as regulatory changes and natural disasters(e.g.,bankruptcy legislation change,Hurricane Katrina).监管变化和自然灾害等外部因素造成的影响（例如，破产法律的变化、卡特里娜飓风）。,Key Factors to Consider in The Consensus Forecast Process共识预测中要考虑的关键因素,6,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,Roll Rate Model 滚动比例模型DMM(Delinquency Migration Matrix)Model DMM（还款逾期移动矩阵）模型Capital Allocation Model 资本分配模型Econometric Model 计量经济学模型Regression Model Using Economic Variables 使用经济变量的回归模型Account Level Models 账户级别模型Logistic Regression on Default Rate(Risk Model)逻辑斯蒂违约率回归（风险模型）Tobit Model on Gross Charge Off Tobit 总撇账额模型Neural Network Model 神经网络模型Other Models Tools Developed By Software Vendors 其他模型 软件公司开发的工具Strategic Analytics Tool 战略分析工具,Loss Forecasting Models and Methodologies损失预测模型和方法,7,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,The roll rate methodology starts with calculating the monthly historical roll rates from current to bucket 1(1 to 29 DPD),from bucket 1 to bucket 2(30 to 59 DPD),etc.,and from 180+DPD to“Charge Off”with available historical observations.Time series models such as simple moving average,weighted moving average,exponential smoothing,and Winters models will be applied to project the future roll rates in the forecasting window,e.g.,roll rates at period(t+1)to(t+12).Forecast for the gross loss is generated according to the outstanding balance forecast and projected roll rates.To account for the seasonality in the roll rate,the seasonal adjusted roll rate model estimates the seasonal factor with historical monthly roll rate data.The final forecast of the roll rate for each bucket is adjusted by the historical seasonal pattern.,Roll Rate Model,Methodology:,Note:DPD represents“Days Past Due”,8,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,The roll rate methodology starts with calculating the monthly historical roll rates from current to bucket 1(1 to 29 DPD),from bucket 1 to bucket 2(30 to 59 DPD),etc.,and from 180+DPD to“Charge Off”with available historical observations.滚动比例法利用现有的历史观察数据，首先计算当期至第1个逾期时间档（逾期天数1至29天）的每月历史滚动比例，然后计算第1逾期时间档至第2逾期时间档（逾期天数30至59天）的每月历史滚动比例，最后计算逾期180天以上到“撇账”的历史滚动比例。Time series models such as simple moving average,weighted moving average,exponential smoothing,and Winters models will be applied to project the future roll rates in the forecasting window,e.g.,roll rates at period(t+1)to(t+12).使用简单移动平均、加权移动平均、指数光华和Winter模型等时间序列模型来预测在预测窗口期内的远期滚动比例，例如期间（t+1）至（t+12）的滚动比例。Forecast for the gross loss is generated according to the outstanding balance forecast and projected roll rates.总损失的预测根据未清偿余额预测和预测滚动比例来计算。To account for the seasonality in the roll rate,the seasonal adjusted roll rate model estimates the seasonal factor with historical monthly roll rate data.为了将季节性因素纳入滚动比例，使用经过季节性调整的滚动比例模型，依据历史上的每月滚动比例时局来估算季节性因素。The final forecast of the roll rate for each bucket is adjusted by the historical seasonal pattern.对每个逾期时间档的滚动比例最终预测要根据历史上的季节性规律进行调整。,滚动比例模型,方法：,逾期时间档,B0,B1,B2,B3,B4,B5,B6,B7,撇账,期间（t-1）,当期,0-29 DPD,30-59 DPD,60-89 DPD,90-119 DPD,120-149 DPD,150-179 DPD,180+DPD,撇账,余额,B0(t-1),B1(t-1),B2(t-1),B3(t-1),B4(T-1),B5(t-1),B6(t-1),B7(t-1),CO(t-1),期间（t）,当期,0-29 DPD,30-59 DPD,60-89 DPD,90-119 DPD,120-149 DPD,150-179 DPD,180+DPD,撇账,余额,B0(t),B1(t),B2(t),B3(t),B4(t),B5(t),B6(t),B7(t),CO(t),滚动比例,当期至 B1,B1至B2,B2至B3,B3至B4,B4至B5,B5至B6,B6至B7,B7至撇账,B1(t)/B0(t-1),B2(t)/B1(t-1),B3(t)/B2(t-1),B4(t)/B3(t-1),B5(t)/B4(t-1),B6(t)/B5(t-1),B7(t)/B6(t-1),CO(t)/B7(t-1),注：DPD 代表“逾期天数”,9,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,The DMM model takes into account the overall delinquency migration of balance,and assumes that the near term delinquency migration will resemble the stable migration matrix estimated.The DMM model starts with calculating the historical Delinquency Migration Matrix from one period to another or from one month to the next month.The stable migration matrix is estimated with a moving average of the historical migration matrix.The charge off for the next period is projected by applying the estimated charge off rate from the migration matrix to the current balance distribution by bucket.,DMM(Delinquency Migration Matrix)Model,Methodology:,Let Bt be the beginning balance at t,M be the delinquency migration matrix,and C be the one period charge off matrix,then,the charge off COt=Bt*C;and the ending balance:Bt1=M*Bt.Thus,the beginning balance at t+1 period can be calculated as Bt+1=M*Bt+B,where B is the adjustments made with new balance for time period t+1.(B could also be estimated with historical data).Therefore,charge off at period t+1 is projected as:COt+1=Bt+1*C.Continue this process to forecast for the p period ahead charge off COt+p with the migration matrix.,B,M,C,A typical delinquency migration matrix is showed as follow:,10,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,DMM 模型考虑了余额的总体逾期还款移动，并假设近期的还款逾期移动所遵守的规律与估算的稳定移动矩阵相似。DMM 模型首先计算从一个期到另一个期间或一个月到下一个月的历史还款移动矩阵。稳定移动矩阵是对历史移动矩阵进行移动平均估算得出的。下一期间的撇账预测方法是，使用从移动矩阵中估算得出撇账率的乘以逾期时间档化的当期余额。,DMM（还款逾期移动矩阵）模型,方法：,设 Bt为 t 时的期初余额，M为还款逾期矩阵，C 为一个期间的撇账矩阵，则撇账 COt=Bt*C；期末余额：Bt1=M*Bt。于是，t+1 期间的期初余额计算公式为 Bt+1=M*Bt+B，式中 B 为对时间期间 t+1 的新余额调整项。（B也可用历史数据进行估算）。因此，期间t+1的撇账预测公式为：COt+1=Bt+1*C。继续这个过程，通过移动矩阵可以预测未来 p 个期间的撇账 COt+p。,B,M,C,以下是一个典型的还款逾期移动模型：,11,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,Capital Allocation Model,Methodology:,Segments,Segments,Historical Loss Rate By Segment,FICO Band,and Vintage,Current Balance Allocation By Segment,FICO Band,and Vintage,Capital allocation model starts with dividing the portfolio into small cells by segment,risk attribute,and vintage,and then estimating the loss rate for the predicted periods based on the corresponding historical loss rates.The projected losses is calculated by the current outstanding balance distribution by segment,risk attribute,and vintage multiplying with corresponding historical loss rate for each cell.Segments could be based on geographic location,market distribution channel,and product etc.,X,i,j,Note:FICO represents“Fair,Isaac&Company”,12,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,资本分配模型,方法：,类别,类别,按类别、FICO分档和账龄统计的历史损失率,按类别、FICO分档和账龄统计的当前余额分配,资本分配模型首先按类别、风险属性和账龄将资产组合分为许多组，然后根据对应的历史损失率估算远期的损失率。预测损失的计算方法是用按类别、风险属性和账龄分配的当前未清偿余额乘以每一组的历史损失率。类别可以地理位置、市场分销渠道和产品等为分类依据。,12个月,12至36,个月,36+个月,B12,B13,B22,B23,B32,B33,B42,B43,B52,B53,账龄,注：FICO代表“Fair,Isaac&Company”,13,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,The regression model estimates quarterly loss rate against the key economic variables that impact the portfolio losses.这个回归模型使用与资产组合损失有关的关键经济变量来估计季度损失率。Key economic variables considered in the regression equation are real Gross Domestic Product(GDP)growth rate,unemployment rate,inflation,prime rate(interest rate),etc.本回归方程中考虑的关键经济变量为实际国民生产总值（GDP）增长速度、失业率、通货膨胀率、基准利率等。The quarterly loss forecast is calculated based on the project quarterly loss rate from the regression equation.根据从回归方程中得到的预测季度损失率来计算季度损失预测。The loss forecast is highly dependent on the outlook of economy forecast on key economic variables.损失预测高度依赖于关键经济变量决定的经济景气度。,Econometric Model(Economic Variables)计量经济学模型（经济变量）,Methodology:方法,A simple econometric loss model is specified as 以下是一个简单的计量经济学损失模型Yt=+*Xt+t Where Yt is the quarterly loss rate.公式中 Yt 为季度损失率。Xt is a vector of economic variables or lag of economic variables.Xt 为经济变量或变量滞后值的向量。and are vectors of parameters estimated from the regression.和 为回归得到的参数估计值向量。t is the error term.t为误差项。,14,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,Account Level Model:Logistic Regression Model账户级模型：逻辑斯蒂回归模型,Methodology:方法,A typical logistic regression on the default rate can be specified as对违约率进行逻辑斯蒂回归可使用下面的常用公式log(Pi/(1-Pi)=+*XiWhere Pi is the default rate in the forecasting window.公式中 Pi 为预测窗口期的违约率。Xi is the vector of explanatory variables(e.g.,individual customer bureau attributes,historical usage pattern and payment behavior etc.)Xi为解释性变量构成的向量（例如，历史使用规律和付款行为等）is the intercept.为截距。is the vector of parameters estimated from the regression.为回归得到的参数估计值向量。,The logistic regression model estimates probability of default for an individual account in the forecasting window.逻辑斯蒂回归模型可估计预测窗口期内具体账户的违约概率。Loss in the forecasting window for an account=Projected Probability of Default*Severity(LGD)for the account.一个账户在预测窗口期内的损失=预测违约概率*账户的严重度（违约损失）。The losses for the portfolio can be calculated by aggregating individual accounts projected loss.具体账户的预计损失加在一起，可计算出信贷组合的损失。,15,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,Account Level Model:Tobit Model账户级别模型：Tobit 模型,Methodology:方法,A typical Tobit(censored regression)model on gross losses can be specified as 以下是一个典型的 Tobit（删失回归）总撇账额模型Yi=*Xi+i if RHS 0(Charge Off Accounts)如果 RHS0（撇账账户）Yi=0 Otherwise(non-Charge Off Accounts)否则（非撇账账户）Where Yi is the gross loss for an account in the forecasting window.公式中Yi为预测窗口期内某个账户的总损失。Xi is the vector of explanatory variables(e.g.,individual customer bureau attributes,historical usage pattern and payment behavior etc.).Xi为解释性变量构成的向量（例如客户个人的征信属性、历史使用规律和付款行为等）。is a vector of parameters estimated from the regression.为回归得到的参数估计值向量。i are residuals that are independently and normally distributed,with mean zero and a common variance 2.i为独立正态分布的参差，平均值为零，共同方差为 2。,Tobit model offers explanatory insight through account level regression that predict the probability of an account charge off and severity of a charge off account.Tobit 模型是通过预测具体账户撇账概率和撇账账户严重性的账户级回归来提供解释性的看法。The gross losses for the portfolio can be calculated by aggregating individual accounts predicted gross loss from the regression equation.资产组合的总损失计算方法是将回归方程得到的具体账户预测总损失累加在一起。,Prob(Yi=0)=1-(*Xi/)(non charge offs,censored)（无撇账，已删失）,16,Confidential and Proprietary Materials.For use only by China Construction Bank and Bank of America Personnel.内部资料，切勿外传。仅供中国建设银行及美国银行人员使用,Hidden Layer 隐蔽层,Input Layer 输入层,Hidden Layer隐蔽层,Output Layer 输出层,Hidden Layer隐蔽层,Output Layer输出层,A typical neural network is depicted in the following diagram:下图所示为一个典型的神经网络：,Account Level Model:Neural Network账户级别模型：神经网络,Methodology:方法,A neural network is a collection of simple computational elements that are interconnected.神经网络是由许多相互连接的简单的计算元素构成的整体。Pi is simply a linear combination of weights for the inputs Xj to the neuron:Pi 就是神经元输入量Xj的权重的线性组合：Pi=Wji*Xj.A transfer function is used by the neuron to converts the sum of weighted inputs(Wji*Xj)into Yi(a value between 0 and 1).神经元采用传递函