FSS Workshop Shadow Bond Approach through Large Corp Scorecard (LCS) Incheol Jo Head, Integrated Risk SC First Bank Contents 1. Overview Of Large Corp Scorecard 2. Model Development Approach 3. Development Process 4. Model Structure of LCS 2 Scorecard Construction EXPERT JUDGEMENT DEVELOPMENT METHODS VALIDATION METHODS •Selection and weighting of factors through experts SHADOW-BOND METHOD GOOD/BAD ANALYSIS •Statistical model built to mimic • Statistical comparison of external ratings, which can defaulted and non-defaulted then be applied to unrated clients to identify factors that companies are predictive of default • Linear regression against (log) PDs of external ratings •Preferably logistic regression, alternatively multivariate discriminance analysis or neural nets •Mapping to external ratings where possible • Measurement of ‘matching statistics’ • Test on hold-out sample •Back-testing against actual default experience/external ratings •Back-testing against actual default experience/external ratings •Alternative: Cross validation •Back-testing against actual default experience POTENTIAL DISCRIMINATORY POWER • Highly dependent on quality of • High discriminatory power expert judgment •Limited by quality of external • Typically not better than rating statistical methods •Highest discriminatory power possible, with a danger of over-fitting DATA AVAILABILITY •Insufficient data to develop a •Insufficient data for good/bad statistical model (few defaults analysis, but external ratings or external ratings available) available for a statistically significant sample of clients •Sufficient data available for at least 100 defaulted and 200 non-defaulted clients 3 LCS at a Glance Customer Segment Description Model Output Model Approval Use Development & Calibration Data Corporate with annual sales turnover above USD 375 million Low default portfolio – Shadow Bond approach using S&P bond ratings as basis for ordinal ranking of risk to select the most predictive financial factors for counter-parties using 20 years of S&P information Scores which then translates into PD, bucketed into CG Approval authority lies with the Wholesale Bank Risk Committee based on recommendation from the Model Assessment Committee Credit Approval, Limit Setting, Credit Risk Reporting, Risk-Based Pricing 252 S&P rated large corporate counterparties in SCB from 20012003. (LCS Version 2) 4 Contents 1. Overview Of LCS 2. Model Development Approach 3. Development Process 4. Model Structure of LCS 5 Model Development Approach 1. Identify appropriate external rating agency as default reference for data sample 2. Specify selection criteria for data sample that reflects SCB’s LCS portfolio 3. Create a list of independent variables 4. Transform each variable 5. Analyze predictiveness and correlation of individual variables 6. Short-list variables and build regression database 7. Run multivariate linear regression 8. Select the best model 9. Test stability 10. Calibrate the model 6 Overview of Data All Large Corporate Customers with External S&P Local Currency Rating Methodology Data, Analysis, Calibration, Re-rating (using recent financials) Benchmarked against S&P bond ratings Removed ratings with parental support, sovereign cap/enhancement 307 data points used for development Data set covering 2000 – 2003 Aim To achieve a good R² model which fits closely to the S&P Local Currency Rating Why S&P rating? To find out rating of a customer’s intrinsic creditworthiness (standalone rating) and S&P is the only international rating agency to produce a local currency rating consistent with SCB’s framework 7 2000 – 2003 S&P Ratings used for Large Corporate Scorecard Development Sample: S&P Rated 307 Developmen t Calibration 252 Cross Validation 55 8 Scorecard Development Data Set No. of Clients by Geographical Region 160 140 131 118 120 100 80 58 60 40 20 0 Europe Americas Asia 9 Scorecard Development Data Set No. of Clients by S&P Rating 60 50 46 41 40 32 29 30 20 36 34 19 16 19 15 7 10 7 3 2 1 0 AAA AA+ AA AA- A+ A A- BBB+ BBB BBB- BB+ BB BB- B+ B- 10 Model Development Approach 1. Identify appropriate external rating agency as reference for model development 2. Specify selection criteria for a data sample that reflects SCB’s LCS portfolio 3. Create a list of independent variables 4. Transform each variable 5. Analyse predictiveness and correlation of individual variables 6. Short-list variables and build regression database 7. Run multivariate linear regression 8. Select the best model 9. Test stability 10. Calibrate the model 11 Data Sample Selection reflecting the Large Corporate Portfolio Selection criteria includes 1. Existence of external ratings 2. Only those externally rated accounts without parental support or sovereign cap/support was admissible to the development/validation sample 3. Sample broadly reflects SCB’s overall portfolio characteristics in terms of industry, size, geography, and type of counterparties 4. Take random sample from (3) for out-of-sample validation 12 Matching SCB customers with S&P rating 13 Model Development Approach 1. Identify appropriate external rating agency as default reference for data sample 2. Specify selection criteria for a data sample that reflects SCB’s LCS portfolio 3. Create a list of independent variables 4. Transform each variable 5. Analyze predictiveness and correlation of individual variables 6. Short-list variables and build regression database 7. Run multivariate linear regression 8. Select the best model 9. Test stability 10. Calibrate the model 14 Financial Factor Selection Long list of ~ 190+ financial factors Statistical analysis of individual factors’ predictive power against external S&P rating Shortlist of highly predictive factors Correlation analysis between factors Statistical analysis to optimise weighting of financial factors Final selection of financial factor sub-model 15 Factor Long List 16 Single Factor Analysis Average Default Rates 100% 90% 80% Portfolio Distribution Average Default Rate ADR + 1 stdev ADR - 1 stdev Score 70% 60% 50% 40% 30% 20% 10% 50 47 45 42 40 37 35 32 30 27 25 22 20 17 15 12 10 7. 5 2. 0 0% 17 Variable Transformation using Logistic Function Transformed Factor Outliers Outliers are are squeezed squeezed towards towards the the centre centre Largest Largest company company concentration concentration isis pulled pulled apart apart Factor Outliers Most companies fall in this range 1 ScX i 1 Expc0 c1 X i Outliers Where Xi stands for the realisation of factor X for client i, and ScXi is the score of factor X for client i. 18 Variable Transformation using Logistic Function - Example Score 1 100 80 0.5 60 0 40 20 -0.5 0 -1 -20 -40 -1.5 -2 Profit007a - Ratio Profit007a - Score -2.5 Score Ratio Ratio -60 -80 -100 19 Model Development Approach 1. Identify appropriate external rating agency as default reference for data sample 2. Specify selection criteria for a data sample that reflects SCB’s LCS portfolio 3. Create a list of independent variables 4. Transform each variable 5. Analyze predictiveness and correlation of individual variables 6. Short-list variables and build regression database 7. Run multivariate linear regression 8. Select the best model 9. Test stability 10. Calibrate the model 20 Short listing Variables Variables would be selected if it exhibits: 1) High power-statistics and R² 2) Good data availability If the ratios satisfy the above criteria, simple ratios are preferred over complicated ones Rank (1-23) Numerator 1 EBITDA 2 Pretax Profit PAT + 3 Depreciation/Amort Other Income 4 Pretax Profit PAT + 5 Depreciation/Amort Other Income 6 EBIT Denominator Interest Charge Total Assets Good. Good Comments Total Liabilities Good Net Sales Good Current Liabilities Total Debt + Equity Total Liabilities - Cash Fixed Deposits Good. 7 Pretax Profit 8 Net Cash After Operations ST Debt + LT Debt OK 9 Total Debt Pretax Profit + Depreciation Extraordinary Items OK 10 11 12 13 14 15 Net Cash After Operations Gross Cashflow less Interest and Dividend Total Equity Total Equity Net Sales Gross Profit (from Trading) 16 Total Asset Growth 17 Total Equity 18 EBIT + Depreciation 19 21 Equity Profit After Tax before Extraordinary Items Current Assets 22 Current Liabilities 23 Gross Profit Growth 20 Total Liabilities Good Total Current Liabilities Good Total Assets -NAInterest Charge Net Sales OK Good. Huge intangibles OK Good, sensitive to M&A activity Trade Creditors May not be very intuitive OK but what happens Current Liabilities -Cash when Cash > Current Liabilities? Sales May not be intuitive -NA- Total Equity Average Total Assets Total Assets -NA- Average & may not be very intuitive Average, sensitive to M&A activity 21 Correlation Analysis -LN(PD) -LN(PD) 100.0% DCS001b_old2_incl amort 47.7% Profit010 41.9% DSC_EBITDA to Debt 44.4% Liq017b 37.1% Profit007 37.4% DCS104 39.2% Profit006 34.5% DCS014_addl1 27.2% DCS103b 29.2% Solv006a_checkif 42.6% DCS018 25.3% DCS014 21.5% Gear005 12.0% Size2 44.9% Profit018 24.7% DCS001b_old2 DSC_EBIT DCS014 Solv006a_c _incl amort Profit010 DA to Debt Liq017b Profit007 DCS104 Profit006 _addl1 DCS103b heckif DCS018 DCS014 Gear005 100.0% 50.3% 77.1% 62.0% 40.9% 51.5% 61.1% 51.0% 58.7% 72.1% 42.8% 32.5% 19.7% 22.1% 17.1% 100.0% 55.9% 68.9% 68.5% 84.7% 76.4% 51.3% 32.9% 56.0% 46.5% 40.2% 25.0% 12.1% 33.2% 100.0% 67.6% 37.0% 65.7% 64.9% 54.7% 74.6% 87.8% 55.8% 35.9% 19.2% 12.2% 11.2% 100.0% 67.7% 56.8% 75.0% 72.4% 51.5% 65.1% 72.2% 59.3% 42.1% 11.8% 34.5% 100.0% 55.6% 63.5% 42.3% 18.5% 39.6% 44.9% 48.5% 34.9% 18.3% 57.8% 100.0% 67.8% 40.8% 36.4% 62.0% 38.3% 28.7% -6.4% 2.7% 23.5% 100.0% 52.9% 45.2% 59.0% 53.9% 39.1% 26.7% 6.1% 26.1% 100.0% 54.4% 52.9% 68.9% 84.5% 34.1% 10.4% 28.3% 100.0% 63.8% 74.1% 37.0% 18.7% 9.0% 1.1% 100.0% 53.8% 40.1% 18.3% 14.0% 20.8% 100.0% 59.2% 31.0% 7.6% 26.2% 100.0% 27.9% 6.9% 35.1% Size2 100.0% 17.8% 100.0% 22.2% 15.3% Profit018 100.0% Variables are selected based on correlation: While certain factors seem to be highly predictive, the use of these factors could sometimes turn out to be less than optimal. This is due to the highly correlated nature with other variables. In these cases, the use of a less predictive, but lowly correlated factor in the final model build may provide a better scoring model. 22 Model Development Approach 1. Identify appropriate external rating agency as default reference for data sample 2. Specify selection criteria for a data sample that reflects SCB’s LCS portfolio 3. Create a list of independent variables 4. Transform each variable 5. Analyze predictiveness and correlation of individual variables 6. Short-list variables and build regression database 7. Run multivariate linear regression 8. Select the best model 9. Test stability 10. Calibrate the model 23 Multi-Factor Analysis Application of multivariate regression to obtain the best combination of factors to reflect the S&P rating (implied S&P PD) Selection of models Statistically optimal models are shortlisted Ascertain whether the models can be improved incorporating credit experts’ recommendation without losing statistical power Choice of qualitative factors is in such a way that there will not be any double counting with financial ratios Stability verified through testing the scorecard on a holdout sample 24 Multi-Factor Analysis PD o data o 100% o o o o o i o o o Score Y 1 1 exp 0 1 X1 + 2 X 2 + 3 X 3 + .... + n X n Statistical Optimisation using maximum log-likelihood function max ln LL max 0 ,..., n 0 ,..., n PD N i 1 observed , i ln PD predicted, i ( 1 PDobserved , i ) ( 1 ln PD predicted, i ) 25 Multi-Factor Analysis DCS001b _old2_incl _amort Profit010 DSC_EBI TDA_to_D ebt Liq017b -0.012417 -0.012417 -0.005857 -0.004264 -0.008212 -0.005914 -0.004794 -0.008399 -0.006248 -0.00795 -0.006248 -0.00795 -0.011076 -0.011076 -0.011076 -0.011076 -0.011076 -0.011076 Profit007 DCS104 DCS014 -0.010058 -0.010058 -0.000299 -0.006466 -0.000756 -0.006134 -0.009989 -0.009989 -0.013349 -0.013349 -0.013349 -0.013349 -0.013349 -0.013349 Size2 -0.020543 -0.020543 -0.019455 -0.019412 -0.019812 -0.019812 -0.020108 -0.020108 -0.020108 -0.020108 -0.020108 -0.020108 Profit018 -0.004821 -0.004821 -0.003407 -0.003722 -0.00432 -0.00432 _LNLIKE_ -528.7525 -528.7525 -526.5282 -526.5776 -527.2535 -527.2535 -530.3955 -530.3955 -530.3955 -530.3955 -530.3955 -530.3955 METHOD BS12 BS10 BS8 BS6 SW20 SW15 SW10 SW05 FS10 FS08 FS06 FS05 Model No. # Factor 2 5 8 11 DCS001b_old2_incl amort 1 Debt Servicing 1 (Int. Coverage) 22.5% 22.5% 22.5% 22.5% Liq017b 2 Liquidity 15.0% 15.0% 15.0% 15.0% Gear005 3 Gearing 15.0% 15.0% 15.0% 15.0% Size2 4 Size 15.0% 15.0% 15.0% 15.0% 17.5% 17.5% NCAO 17.5% 17.5% RoA (or) 15.0% DCS014_addl1 5 DCS018 Profit010 Profit007 6 Gross Cashflow (or) NPM 15.0% 15.0% TOTAL 100.0% 100.0% 15.0% 100.0% 100.0% 26 Discriminatory Power of Model Measured by Adjusted Power Statistic Cumulative Number of Implied PD Perfect discriminatory power (“Crystal Ball”) Total Implied PD Power Statistic = Highest discriminatory power: PS = 100% Lowest discriminatory power: PS = 0% B A Realistic model A A+B No discriminator y power Total Customers Number of customers (ordered from worst to best) 27 Model Development Approach 1. Identify appropriate external rating agency as default reference for data sample 2. Specify selection criteria for a data sample that reflects SCB’s LCS portfolio 3. Create a list of independent variables 4. Transform each variable 5. Analyse predictiveness and correlation of individual variables 6. Short-list variables and build regression database 7. Run multivariate linear regression 8. Select the best model 9. Test stability 10. Calibrate the model 28 Model Calibration Continuous function transforms the scores into a PD which in turn is bucketed into different credit grades based on SCB’s masterscale 1 PD = {1 + e (Co-efficient * Score + Intercept)} Co-efficient Indicates correlation of the score with the probability of default Score Overall score which incorporates the aggregated financial and non-financial scores Intercept Value / Intercept used as a factor of transformation / anchor for the logistic function. 29 Model Calibration ► Scores are transformed into a PD using a mathematical function ► PDs are then bucketed into CGs using the PD ranges in Expanded Credit Master Scale (ECMS) Illustrative Calibration Approach % Customers in Portfolio Calibration Curve PD 8% 25% 8.00% PD Calibration Curve 7.00% 20% 6% 6.00% PD% 7% 5.00% 5% 15% 4.00% 4% 3.00% 10% 3% 2.00% 2% 1.00% 5% 1% 0.00% 100 90 80 70 60 50 Score 40 30 20 10 0 0% 0% 1A 1B 2A 2B 3A 3B 4A 4B 5A 5B6A 6B7A 7B 8A 8B9A 9B10A10B10C11A11B11C Credit Risk Grades 30 Contents 1. Overview Of LCS 2. Model Development Approach 3. Development Process 4. Model Structure of LCS 31 Scorecard Architecture Mapping PD into CRG Financial Factors (e.g. size) Non-Financial Factors (e.g. industry) “Warning Signals” (e.g. fraud, Stale Financials) Financial Score Non-Financial Score Warning Signal Grade Change Size Cap/ adjustments Standalone Rating Parental Support/Cap Supported Rating Sovereign Ceiling Final Rating Override 32 Financial Model Overall Weight for Financials : 80% Total Equity Profit Before Tax Net Sales Total Tangible Assets CAPITAL STRUCTURE Weight: 15% EBITDA Interest Expense DEBT SERVICING Weight: 22.5% LIQUIDITY Profit After Tax+Dep/AmorOther income (Non-cash) Total Debt Weight: 7.5% PROFITABILITY Weight: 17.5% Financial Model Net Cash After Ops Total Liabilities NET CASH FLOW Weight: 20% Total Equity SIZE Weight: 17.5% 33 Non-Financial Model Overall Weight for Non-Financials : 20% Industry Risk Weight: 60% Qualitative Model Board/CEO Governance Weight: 40% 34 Warning Signal S. No. 1 2 3 4 5 6 7 8 9 10 Warning Signal Type Request for interest/repayment suspense Internal incidence of fraud at the board level Breach of covenants (analyst needs to specify the type of covenant) Internal incidence of fraud below board level External incidence of fraud Unforced changes of auditor (for the worse) If the auditor is not on the list of SCB-recognized auditors If the financials do not reflect the true creditworthiness of the borrower (e.g. “window dressing”, unconsolidated financials) If the financials are qualified or company prepared Timeliness of Financials Impact 1 - 3 notches/ CRG 12 1 – 3 notches 0 – 2 notches 0 – 1 notch 0 – 1 notch 0 – 1 notch 0 – 1 notch 0 – 2 notches 0 – 1 notch 0 – 1 notch Warning signals: Notching down mechanism For example: For a standalone rating of 7A, if signal 5 (External incidence of fraud) is selected, then user can decide to notch down by 0 or 1 notch, to either 7A or 8A. For a standalone rating of 7B, if signal 5 (External incidence of fraud) is selected, then user can decide to notch down by 0 or 1 notch, to either 7B or 8B. 35 Scorecard Architecture Parental Support Support Level Unqualified Contractual - Unconditional legal obligation Unqualified Non Contractual - Standard Letter of Comfort / Awareness - Non-standard Letter of Comfort / Awareness - Comprehensive verbal statement of support (needs majority ownership) Qualified Contractual - Conditional legal obligation Qualified Non Contractual - Non-Comprehensive Verbal Support Support Mechanics “P” “P-1/P-2” “P-1” “S” to “S+3” (no better than “P- 2”) Eligibility Criteria for Parental Support: • The borrower has to be of strategic importance to the parent company (e.g similar branding, same name, etc) • The borrower has to be of significant economic importance to the parent company • If the parent and subsidiary operates in different countries, country rating of parent company has to be CG 6B or better 36