Best practice in data & scoring Dr Paul Russell Director Analytical Solutions © Experian Limited 2007. All rights reserved. Experian and the marks used herein are service marks or registered trademarks of Experian Limited. Other product and company names mentioned herein may be the trademarks of their respective owners. No part of this copyrighted work may be reproduced, modified, or distributed in any form or manner without the prior written permission of Experian Limited. Confidential and proprietary. Agenda Some themes Analytics and the customer life cycle The role of scoring Building a scorecard Using scoring systems Risk management infrastructure © Experian Limited 2007. All rights reserved. Confidential and proprietary. 2 Themes Best practice is often discussed but almost never seen Do the simple things well Risk management is more than just a scorecard The same principles apply across the credit lifecycle © Experian Limited 2007. All rights reserved. Confidential and proprietary. 3 13 ways to grow bad debt Credit process Target step population Customer acquisition New customers Description 1. Identifying potential customers; 2. Selling credit products to new customers; 3. Identifying the credit risk of the customer and the proposed transaction; 4. Identifying the risk of fraudulent application 5. Deciding whether to accept or decline the transaction; Customer management Collections Existing, nondelinquent customers Existing, delinquent customers 6. Deciding, for accepted transactions, on the terms, e.g., credit amount, pricing. 7. Reviewing the customers facilities (e.g., credit limits, price, etc.); 8. Cross-selling new products to the customers; 9. Ensuring good customers are retained; 10. Identify fraudulent transactions. 11. Identifying self-cure customers; 12. Rehabilitation of potentially good customers; 13. Work-out customers where relationship is broken. © Experian Limited 2007. All rights reserved. Confidential and proprietary. 4 Why is credit risk management important? 2,986 European consumer finance business, Profit Before Tax and Impairment Charges ($m) Get it right and it can support phenomenal value creation 2,196 1,522 1,520 1,382 Profits 1,374 Impairment charges 1,230 1,094 924 836 802 740 676 804 764 478 288 1998 340 1999 2000 2001 2002 2003 © Experian Limited 2007. All rights reserved. Confidential and proprietary. Source: Annual Reports 2004 2005 2006 5 5 core components Component Description Data Application data (for new customers) Account behaviour data (for existing customers) External data (e.g., credit bureaux) Statistical Models Risk models (PD, LGD), fraud models (application and transaction fraud) and revenue models Credit strategies Business rules that translate the outcome of statistical models in credit decisions (accept/decline, price, credit limits, etc.) that maximise profit Implementation tools Software tools to automate the calculation of the above scores and credit strategies on-line on high volumes, with a high degree of flexibility to change credit strategies “on the fly” Evaluation tools Software tools to evaluate the performance of statistical models and credit strategies, and accuracy of implementation © Experian Limited 2007. All rights reserved. Confidential and proprietary. 6 Agenda Some basic themes Analytics and the customer life cycle The role of scoring Building a scorecard Using scoring systems Risk management infrastructure © Experian Limited 2007. All rights reserved. Confidential and proprietary. 7 Analytics and the customer life cycle Solicitation Application Customer management Collections Debt recovery Population Information Analytics touches every part of the customer lifecycle Analytics touches every part of the customer life cycle Amount of information about the customer grows as the relationship advances through the customer life cycle © Experian Limited 2007. All rights reserved. Confidential and proprietary. 8 Analytics and the customer life cycle Solicitation • Channel preference • Contact history • Demographics • Location • Bureau data • Action outcomes • Costs Application • Channel • Product holdings • Demographics • Bureau data • Previous relationships • Account performance • Costs Customer management Collections Debt recovery • Product holdings • Usage • Delinquency • Customer contacts • Preferences • Bureau data • Actions taken • Action outcomes • Costs • Action history • Promises to pay • Promises fulfilled • Action outcomes • Bureau data • Costs • Action history • Promises to pay • Bureau data • Agents used • Promises fulfilled • Litigation outcomes • Costs © Experian Limited 2007. All rights reserved. Confidential and proprietary. 9 Analytics and the customer life cycle Define Goals Agree objectives Understand results Plan Assess current challenger Review Assess Strategy Review Design Monitor Track progress against expectations Implement Build new strategy Ensure operational deployment © Experian Limited 2007. All rights reserved. Confidential and proprietary. 10 Agenda Some basic themes Analytics and the customer life cycle The role of scoring Building a scorecard Using scoring systems Risk management infrastructure © Experian Limited 2007. All rights reserved. Confidential and proprietary. 11 The role of scoring Credit scoring is a technique for predicting the future This prediction can be anything of importance to the business Arrears Fraud Profit Response Account closure Company failure Etc. All scoring is based on one key assumption: The past predicts the future © Experian Limited 2007. All rights reserved. Confidential and proprietary. 12 The role of scoring Example Scorecard How does scoring work? Baseline Constant • Scorecards add and subtract points to a baseline constant according to individual’s or account’s data Applicant Age in Years < 22 -50 22 - 25 -20 26 – 40 0 41 – 55 +30 > 55 • Scorecards are easy to apply and simple to understand The resulting score gives a prediction of future behaviour 0 Worst Status L6M (on all Accounts) 0 • 800 0 1-2 -45 3+ -100 Joint Applicant Present • Scores are used to rank individuals to assign the best actions Y +20 N 0 Etc. … © Experian Limited 2007. All rights reserved. Confidential and proprietary. Etc. … 13 The role of scoring – application scorecard • • Consider a scorecard built to predict whether a new applicant for a credit product will default in the next 12 months This scorecard is used when a new customer applies… Scorecard Application Form Data Score External Data (Bureau etc.) Take most appropriate action for each individual © Experian Limited 2007. All rights reserved. Confidential and proprietary. 14 Proportion of Applicants The role of scoring - scores can drive actions Application Score Low Score / High Risk High Score / Low Risk Extremely Low Risk Extremely High Risk Reject High Risk Standard Risk Reject or price to cover the high expected loss Accept on standard terms © Experian Limited 2007. All rights reserved. Confidential and proprietary. Consider for crosssell of other products 15 The role of scoring - benefits Best use of data Objective Consistent Automation Control Reduced losses © Experian Limited 2007. All rights reserved. Confidential and proprietary. 16 Agenda Some basic themes Analytics and the customer life cycle The role of scoring Building a scorecard Using scoring systems Risk management infrastructure © Experian Limited 2007. All rights reserved. Confidential and proprietary. 17 Building a scorecard – 3 requirements Development sample – the historical data on which the scorecard will be built Outcome – what we are trying to predict Some time later NOW The recent past THEN Modelling methodology – the statistical tool that will help us form our scoring model Outcome Development Sample Statistical Model © Experian Limited 2007. All rights reserved. Confidential and proprietary. Scorecard 18 Representative Products Business cycle The future THEN Is my sample any good? The recent past Development Sample Robust Volumes Mature Is the outcome reliable? © Experian Limited 2007. All rights reserved. Confidential and proprietary. 19 • • This data can come from a number of sources All relevant data should go into the development sample THEN Building a scorecard – the development sample The recent past Development Sample Application Form Credit Bureau Data Historical Account Behaviour Other Account Information Information collected from the applicant at the application point Information on the individual’s other credit commitments Information on the historical behaviour on the account Information on the historical behaviour on other accounts with the same lender © Experian Limited 2007. All rights reserved. Confidential and proprietary. 20 Building a scorecard – the outcome • • Can be a continuous variable (profit, revenue, loss given default, etc.) More commonly it is dichotomous - yes/no Will this applicant default? Is this transaction fraudulent? Will this company fail? Etc. NOW This is the behaviour that we are trying to predict Outcome Good THE FUTURE Bad Observation - Now Outcome - Prediction © Experian Limited 2007. All rights reserved. Confidential and proprietary. 21 NOW What are we trying to predict? Consumer Limited business Non-limited business Bad Good 3 payments in arrears Not 3 payments in arrears Failed Still going Bankruptcy, court judgements or defaults No bankruptcy, court judgements or defaults © Experian Limited 2007. All rights reserved. Confidential and proprietary. Outcome 22 Building a scorecard – the statistical model Many statistical tools available Statistical Model Data is the most important factor Observation Data Outcome Statistical Model Scorecard Statistical tool needs to be: Powerful – to get the best prediction from the data Flexible – can handle varying data types and outcomes Interpretable – easy to understand and to overlay business intelligence Transparent – should be non-’black box’ for regulatory reasons and to ensure understanding © Experian Limited 2007. All rights reserved. Confidential and proprietary. 23 Building a scorecard - the statistical model Statistical Model Linear regression Reality Logistic regression Artificial neural networks Etc Other things being equal the choice of algorithm has relatively little impact on the ultimate power of the model x x x x x x x x x x x x x x x x x x x x x Prediction © Experian Limited 2007. All rights reserved. Confidential and proprietary. 24 Building a scorecard – assessing the model Does the model solve the business problem? Statistical Model Discrimination – the power to polarise individuals between good and bad - Gini statistic & Kolmogorov-Smirnov statistic Accuracy – how much of the variability of the outcome is explained by the model Validation – ensures that over-modelling has not occurred or that an anomalous sample has not been used Improvement – the new model should outperforms the existing model © Experian Limited 2007. All rights reserved. Confidential and proprietary. 25 Agenda Some basic themes Analytics and the customer life cycle The role of scoring Building a scorecard Using scoring systems Risk management infrastructure © Experian Limited 2007. All rights reserved. Confidential and proprietary. 26 Using scoring systems SCORE CARD DATA DECISIONS STRATEGY The data feeds the scoring system, which is used to aid the decisioning The decisions a company makes determine its strategy It is the aims and strategy of the business that must be considered when deciding how to use a scoring system, e.g. Growing the market share Reducing bad debt Increasing automation Maximising response for given marketing cost Combating fraud © Experian Limited 2007. All rights reserved. Confidential and proprietary. 27 Proportion of Applicants The role of scoring - scores can drive actions Application Score Low Score / High Risk High Score / Low Risk Extremely Low Risk Extremely High Risk Reject High Risk Standard Risk Reject or price to cover the high expected loss Accept on standard terms © Experian Limited 2007. All rights reserved. Confidential and proprietary. Consider for crosssell of other products 28 Using scoring systems - the score distribution • • Score Band # Goods # Bads GB Odds % Applicants ≤ 400 500 500 1 9.8 401 – 550 700 350 2 10.3 551 – 650 815 163 5 9.6 651 – 700 1008 84 12 10.7 701 – 750 976 61 16 10.1 751 - 800 950 38 25 9.7 801 – 850 1000 25 40 10.0 851 – 900 1050 21 50 10.5 901 – 950 960 16 60 9.5 ≥ 951 1000 10 100 9.9 TOTAL 8959 1268 7.1 100 Score distribution is obtained by applying the score to the development sample Gives us a prediction for new applicants falling into a given score range © Experian Limited 2007. All rights reserved. Confidential and proprietary. 29 Building a scorecard - the score distribution Score Band # Goods # Bads GB Odds % Applicants ≤ 400 500 500 1 9.8 401 – 550 700 350 2 10.3 551 – 650 815 163 5 9.6 651 – 700 1008 84 12 10.7 701 – 750 976 61 16 10.1 751 - 800 950 38 25 9.7 801 – 850 1000 25 40 10.0 851 – 900 1050 21 50 10.5 901 – 950 960 16 60 9.5 ≥ 951 1000 10 100 9.9 TOTAL 8959 1268 7.1 100 REJECT REFER ACCEPT ACCEPT WITH XSELL Score + Policy Rules + Terms of Business = Strategy © Experian Limited 2007. All rights reserved. Confidential and proprietary. 30 Agenda Some basic themes Analytics and the customer life cycle The role of scoring Building a scorecard Using scoring systems Risk management infrastructure © Experian Limited 2007. All rights reserved. Confidential and proprietary. 31 Implementation – the Business Rules Engine Data Rules execution (Decision Agent) Rules Definition (Strategy Design Studio) Results © Experian Limited 2007. All rights reserved. Confidential and proprietary. Deployed in: Origination Application processing Portfolio Management Customer level decisioning Collections Authorisations Intelligent Messaging Event Management Basel II Stress testing ….. 32 The unsecured lending origination process Gather & validate application data Gather existing customer information A full range of client options and interfaces for channel independence and data accuracy Invoke enrichment strategy Credit bureau links Online links to gather data about existing relationships and customer behaviour Get decision & terms of business Business-driven scoring & decisionmaking Detect application fraud Application screening and data matching Handle referrals and manual procedures Comprehensive workflow capabilities and provision of relevant data for users Get policy decision & enrichment strategy Business-driven scoring and decision-making Implement final decision Automated account set-up. Provision of hand-off files. Letter and e-mail production © Experian Limited 2007. All rights reserved. Confidential and proprietary. 33 Defines Business logic, Segmentation, Scorecards, Strategies and Champion Challenger Business logic, Segmentation, Scorecards, Strategies and Champion Challenger Active History HOST e.g. Account Management System, Authorisation System etc Implements Results Analytical Data Mart Extract Feedback Variables Decision Engine Strategy Implementation Data Manager Operational environment Evaluation Optimisation Reporting © Experian Limited 2007. All rights reserved. Confidential and proprietary. Rule Definition Strategic business environment 34 Beyond scoring - strategy optimisation There are disadvantages to traditional champion/challenger testing… • The time frame for observing results can be long • It can be hard to design the next step • The result can become a “semi-random walk”... We want to get there with the first challenger ! Value Challenger n Using performance data enables better decisions, but is also more complex to combine all the decision influences to maximise value Decision strategy “deploy-learndeploy” process The challenger strategy proven in one time period, may no longer be appropriate for another time period – things change Challenger 1 Champion Challenger 3 Challenger 2 Challenger 4 Time © Experian Limited 2007. All rights reserved. Confidential and proprietary. Influence due to: • Macro-economics? • Use of intuition? • Misunderstanding? 35 #35 Developments in analytics - strategy optimisation The next step… Incremental benefit ROI Some organisations are still here Scoring X XXX X XX XXX X X XX Manual X X X X X X X XX X X X X X XX X X X XX X • Experience and intuition • Trial and error Optimised Strategies Most are here • • Elaborate Strategies • Allocates optimal action for each customer within constraints • Segmentation based on predictive model dimensions: e.g. risk and revenue • Objective, mathematical goal maximisation • “Subjective” judgment used to manage trade-offs XX X X X X X X single predictive model e.g. credit risk score “Heuristic” cut-offs assigned using good:bad odds Underlying decision complexity © Experian Limited 2007. All rights reserved. Confidential and proprietary. 36 #36 Stage 1: build the infrastructure Centralisation of credit decisioning Set-up of IT tools required to automate credit risk and market management processes and the interaction between front line and back office Development of decision support tools Development of credit / marketing databases Automate the processes © Experian Limited 2007. All rights reserved. Confidential and proprietary. 37 Stage 2: fine-tune for performance Fine customer segmentation based on customer profile, product holding and behavior data Advanced credit and marketing databases drive increased sophistication in statistical models development Customer interactions for risk and marketing are proactively initiated at all key points Strategies are designed at customer-level Automate the decisions © Experian Limited 2007. All rights reserved. Confidential and proprietary. 38 Stage 3: optimize for excellence Infrastructure enables total proactive control of the business – decision analytics becomes a way of life Risk and marketing strategies are centrally designed based on advanced statistical techniques and drive customer profitability Decision analytics is well structured and integrated across business functions including risk, marketing, sales, operations, finance Optimize the decisions © Experian Limited 2007. All rights reserved. Confidential and proprietary. 39 The Road Map Build the infrastructure Strategy Fine-tune for performance Profit driven credit strategy in place and reviewed regularly Credit policy in place How are credit policies and strategies defined, reviewed and improved? Processes Include all available data into the process. Focus underwriter on “key” review, not second scorecard How well do staff understand all profit drivers? What is the degree of expertise in credit scoring and decision science? Knowledge Monthly review of credit strategies. Champion/challenger a way of life Processes regularly reviewed and refined. Little manual intervention Processes well defined and automated How well defined and the processes, and what is the degree of automation? Optimize for excellence Tools “a world-class consumer finance company” Processes fully support profit-driven strategy, and are integrated across functions Education on strategy review process, fully understanding the use of MIS Ongoing knowledge improvement Full suite of scorecards, ability to optimize credit strategies Ability to review and modify credit strategies ‘on the fly’ Profit-driven organisation across functions Create strategy review cross-functional team Scorecards in place for all critical segments, decision engine used to Ensure clear assignment of control terms of responsibilities for risk business. Generate management functions key KPI’s What credit management tools are used? How How are credit risk, flexible are they? How marketing and finance easy is it for business user working together? How to change processes and Organisation are operational and strategies? strategic decisions taken? © Experian Limited 2007. All rights reserved. Confidential and proprietary. 40 Conclusions It all starts with data Scorecards are important Strategy is more important Implementing the strategy properly is vital If you don’t monitor you’re wasting you time Risk management is a never-ending journey © Experian Limited 2007. All rights reserved. Confidential and proprietary. 41 Best practice in data & scoring Dr Paul Russell Director Analytical Solutions © Experian Limited 2007. All rights reserved. Experian and the marks used herein are service marks or registered trademarks of Experian Limited. Other product and company names mentioned herein may be the trademarks of their respective owners. No part of this copyrighted work may be reproduced, modified, or distributed in any form or manner without the prior written permission of Experian Limited. Confidential and proprietary.