INTRODUCTION TO RATING MODELS

Transcription

1

2 INTRODUCTION TO RATING MODELS Dr. Daniel Straumann Credit Suisse Credit Portfolio Analytics Zurich, May 26, 2005 May 26, 2005 / Daniel Straumann Slide 2

3 Motivation Due to the Basle II Accord every bank has to assess the default probability (DP) of each obligor (firm, private client, consumer) in its credit portfolio (IRB Approach). This calls for internal rating systems. A rating system is a rational and replicable procedure, which assigns to each obligor a rating (one year DP) based on the following information: market information (equity prices, interest rates, credit spreads, derivatives prices, credit agencies ratings, macroeconomic variables,... ), balance sheet information resp. income data and assets (for individuals), qualitative data (marital status, domicile, industry sector,... ), qualitative assessment (by credit analysts), client s individual financial behavior,... May 26, 2005 / Daniel Straumann Slide 3

4 Examples of Ratings Standard & Poor s Moody s Rating DP Rating DP AAA % Aaa % AA % Aa % AA 0.01 % Aa % AA % Aa % A % A % A 0.05 % A % A % A % BBB % Baa % BBB 0.25 % Baa % BBB % Baa % BB % Ba % BB 1.28 % Ba % BB % Ba % B % B % B 6.45% B % B % B % CCC % Caa & < % Table 1: Masterscales May 26, 2005 / Daniel Straumann Slide 4

5 Brief History of Rating Systems Starts with scorecard models, which are used for evaluating creditworthiness of customers or firms. First proposals during Second World War. Example: Consumer loans. Residential Status Age Loan Purpose Owner New car 41 Tenant Second hand car 33 Living with parents Home improvement 36 Other specified Holiday 19 No response Other 25 Table 2: Simple scorecard for consumer credit. This means: a 20 year old person, living with its parents, wishing to borrow money for a second hand car, has a score of = 69. May 26, 2005 / Daniel Straumann Slide 5

6 Example: Altman s Z score (1968) Z = X X X X X 5, (1) where X 1 = Working capital/total assets, X 2 = Retained earnings/total assets, X 3 = Earnings before interest and taxes/total assets, X 4 = Market value of equity/book value of total debt, X 5 = Sales/Total assets. Classification rule: Z 1.81 bankruptcy, 1.81 < Z 2.99 zone of ignorance, Z > 2.99 non bankruptcy. May 26, 2005 / Daniel Straumann Slide 6

7 Remarks: Coefficients are optimized for a particular data sample. Nevertheless Altman s Z score is still used (can e.g. be downloaded from Bloomberg). Scorecard models are often applications of statistical classification methods. Beginning of the 90 s: Focus on default probabilities. Rating systems in practice RiskCalc (Moody s): Scorecard model CreditEdge (KMV): Structural default model... May 26, 2005 / Daniel Straumann Slide 7

8 The Classification Problem Group of non defaulters (Y = 0) and defaulters (Y = 1). Each obligor has certain (observed) random attributes X R m. We assume: non defaulters have density f 0 (x) = f(x Y = 0), defaulters have density f 1 (x) = f(x Y = 1). Let p = P(Y = 1) = 1 P(Y = 0) be the default probability in the combined sample. Assume classification rule of the following kind for a subset A R m : Ŷ (x) = 1 x A. Classification errors: α = P(Ŷ = 0 Y = 1) = β = P(Ŷ = 1 Y = 0) = R m \A A f 1 (x)dx f 0 (x)dx (1-(hit rate)), (false alarm rate). May 26, 2005 / Daniel Straumann Slide 8

9 Total Misclassification Error TME = P(classification error) = pα + (1 p)β = pf 1 (x)dx + (1 p)f 0 (x)dx. R m \A Aim: Minimize TME. Since R m f 1 (x)dx = 1, we have that pf 1 (x)dx = p R m \A A A pf 1 (x)dx, and can therefore write TME = p + A ((1 p)f 0 (x) pf 1 (x)) dx. Hence TME is minimized for x A (1 p)f 0 (x) pf 1 (x) < 0 S(x) = f 0(x) f 1 (x) < p 1 p. (2) May 26, 2005 / Daniel Straumann Slide 9

10 Remarks We can regard S(x) as a score. Costs of misclassification can easily be incorporated. Minimal TME Rule (2) is equivalent to x A P(Y = 0 x) < P(Y = 1 x). Examples Assume that X N (µ i, Σ i ), i = 0, 1. Then Σ 0 = Σ 1 : x A log S(x) = (µ 0 µ 1 ) T Σ 1 0 (x 0.5(µ 0 + µ 1 )) < log The score log S(x) is a linear function. ( ) p. 1 p May 26, 2005 / Daniel Straumann Slide 10

11 Σ 0 Σ 1 : x A log S(x) = 1 ( ) 2 log det(σ0 ) det(σ 1 ) 1 2 µt 0 Σ 1 0 µ µt 1 Σ 1 1 µt xt (Σ 1 0 Σ 1 1 )x 2xT (Σ 1 0 µ 0 Σ 1 1 µ 1) < log The score log S(x) is a quadratic function. Alternative classification methods Classification trees Nearest neighbor methods Neural networks Expert systems ( p 1 p ). May 26, 2005 / Daniel Straumann Slide 11

12 Logistic Regression For the derivation, assume that X N (µ i, Σ), i = 0, 1. Then ( ) P(Y = 1 X = x) log( posterior odds ) = log = log P(Y = 0 X = x) Definition = const. + α T x, α R m. ( f1 (x)p ) f 0 (x)(1 p) Let Y i be 0 1 response variables and x i R m vectors of explanatory variables. Then we say that Y i obeys a logistic regression if: 1. The Y i s given x i are conditionally independent. 2. There are α R m and c R with where F (u) = 1/(1 + exp( u)). P(Y i = 1 x i ) = F (c + α T x i ), (3) May 26, 2005 / Daniel Straumann Slide 12

13 Remarks The above model is a so called generalized linear model with link function F 1. The choice F (u) = Φ(u) = (2π) 1/2 u exp( s2 /2)ds would correspond to a probit model. The logistic regression is the workhorse for the development of new rating models. Other areas of application: medical and natural sciences. Practical Issues of Calibration Estimation by MLE. Data cleaning, construction of meaningful financial ratios. Variable selection methods (Stepwise forward, stepwise backward). Goodness of fit checks. May 26, 2005 / Daniel Straumann Slide 13

14 Validation of Rating Models Suppose we have constructed a score S = S(X), where X is an observed random vector. We want to predict Y in the following way: Ŷ (x) = 1 S(x) s. (4) The number s R is called cut off point. Strong Discrimination Weak Discrimination Density f1 f0 Density f1 f Score Score We need to measure the distance between f 0 and f 1 in a meaningful way. May 26, 2005 / Daniel Straumann Slide 14

15 The misclassification errors for (4) are given by α(s) = P(S > s Y = 1) (1-hit rate), β(s) = P(S s Y = 0) (false alarm rate). The curve (β(s), 1 α(s)) is called receiver operating characteristics (ROC Curve). ROC Curve Hit Rate False Alarm Rate May 26, 2005 / Daniel Straumann Slide 15

16 AUROC The area under the ROC Curve (AUROC) serves as a measure for the discriminatory power of the score S. Properties of AUROC 0 AUROC 1. Perfect discrimination: AUROC = 1 There is c R such that: Y = 1 S c, AUROC = 0 There is c R such that: Y = 0 S c. If Y and S are independent, then AUROC = 1/2. AUROC is invariant under strictly increasing transforms of S. May 26, 2005 / Daniel Straumann Slide 16

17 AUROC=1 AUROC=0.5 AUROC=0 Default Default Default Score Score Score Figure 1: Representation of joint distributions for AUROC= 1, 0.5, 0. May 26, 2005 / Daniel Straumann Slide 17

18 Practical Applications of AUROC AUROC values of 0.8 and more are often considered as good. Similar to linear correlation, AUROC values are per se not very meaningful and rather difficult to interpret. Danger of misuse: AUROC is not a measure of goodness of fit: add a constant to the logistic regression score, and the AUROC will not change. By the Neyman Pearson lemma, AUROC is bounded from above by a functional depending on the densities f 0 (x) and f 1 (x). May 26, 2005 / Daniel Straumann Slide 18