© Deloitte Consulting, 2004 Alternatives to Credit Scoring in Insurance James Guszcza, FCAS, MAAA Cheng-Sheng Peter Wu, FCAS, ASA, MAAA CAS 2004 Ratemaking Seminar Philadelphia March 12-13, 2004 © Deloitte Consulting, 2004 Agenda Introduction The credit scoring revolution From credit scoring to predictive modeling The big idea: credit scoring is just one kind of insurance predictive (“data mining”) model… many other predictive models can be built Conclusions What it means to actuaries 2 © Deloitte Consulting, 2004 Introduction © Deloitte Consulting, 2004 Introduction – Our Efforts Is credit for real? Going beyond credit “Does Credit Score Really Explain Insurance Losses? Multivariate Analysis from a Data Mining Point of View” “Mining the Most from Credit and Non-Credit Data” How do credit scoring models work? “A View Inside the “Black Box”: Review and Analysis of Personal Lines Insurance Credit Scoring Models Filed in the State of VA” 4 © Deloitte Consulting, 2004 Introduction Which Company is this? $10 Premium (in Billions) $9 $8 $7 $6 $5 $4 $3 $2 $1 $0 1994 1995 1996 1997 1998 1999 2000 2001 2002 Year 5 © Deloitte Consulting, 2004 Introduction 110% 35% 105% 30% 100% 25% 95% 20% 90% 15% 85% 10% 80% 5% 75% 0% 1994 1995 1996 1997 1998 Year 1999 2000 2001 Progressive Combined Ratio Progerssive Growth Rate Premium Growth Rate Combined Ratio Progressive Insurance Company 2002 6 © Deloitte Consulting, 2004 Introduction Progressive Combined Ratio Progressive vs. Industry Industry Combined Ratio Progerssive Growth Rate 115% 35% 110% 30% 105% 25% 100% 20% 95% 15% 90% 10% 85% 5% 80% 0% 1994 1995 1996 1997 1998 Year 1999 2000 2001 2002 Growth Rate Combined Ratio Industry Growth Rate 7 © Deloitte Consulting, 2004 Why? Multiple Choice Progressive provided foosball tables and free snacks to their trendy, 20-something workforce Progressive built a compound GammaPoisson GLM model to design their class plan Progressive pioneered the use of credit in pricing/underwriting 8 © Deloitte Consulting, 2004 The Credit Score Revolution © Deloitte Consulting, 2004 Personal Lines Pricing and Class Plans – History Few rating factors before World War II Explosion of class plan factors after the War Auto class plans: Homeowners class plans: Territory, driver, vehicle, coverage, loss and violation, others, tiers/company… Territory, construction class, protection class, coverage, prior loss, others, tiers/company... Credit scoring introduced in late 80s and early 90s 10 © Deloitte Consulting, 2004 Personal Lines Credit Scoring – History First important factor identified over the past 2 decades Composite multivariate score vs. raw credit information Introduced in late 80s and early 90s Viewed at first as a “secret weapon” Quiet, confidential, controversial, black box, …etc “Early believers and users have gained significant competitive advantage!” 11 © Deloitte Consulting, 2004 The Current Environment Now everyone is using it: Marketing and direct solicitation New business and renewal business pricing and underwriting How to stay competitive if everyone is using it? Regulatory constraints: Many states have conducted studies on the true correlation with loss ratio and potential discrimination issues - WA study, TX study, MO study Many states have/are considering restricting the use of credit scores or certain types of credit information More states want the “black box” filed and opened 12 © Deloitte Consulting, 2004 Some Facts About Credit Scores A composite score that usually contains 10 to 40 pieces of credit information Loss ratio lift is significant – a powerful class plan factor or rate tiering factor Benefits/ROI are measurable Payment pattern information, bankruptcies/liens, collections, inquiries, bad debt/defaults… Lift curve can be translated into bottom-line benefit Blind test and independent validation can be done to verify the benefit 13 © Deloitte Consulting, 2004 Loss Ratio Lift Curve 120 90 82 78 Loss Ratio 74 66 70 62 58 50 Credit Score Decile 14 © Deloitte Consulting, 2004 Credit Score Revolution Segmentation Power 1997 NCCI/Tillinghast Study of 9 Companies' Data Loss Ratio Relativity of the Best and Worst 20% of Credit Score Co1 Co2 Co3 Co4 Co5 Co6 Co7 Co8 Co9 Avg Best 20% -38% -29% -19% -15% -14% -34% -22% -22% -36% -25% Worst 20% 48% 20% 32% 30% 46% 59% 20% 22% 95% 41% 15 © Deloitte Consulting, 2004 From Credit Scoring to Predictive Modeling © Deloitte Consulting, 2004 From Credit Scores to Predictive Models What is a predictive model? A multivariate scoring formula (linear or nonlinear) that combines the values of several predictive variables to estimate the value of a target variable What is a credit score? A multivariate scoring formula (linear or nonlinear) that combines the values of several credit variables to estimate the value of a target variable 17 © Deloitte Consulting, 2004 From Credit Scores to Predictive Models A credit score is just one example of an insurance predictive model A credit scoring project is a first approximation to a full insurance data mining project. The same methods used to build credit scores are used in data mining to build insurance predictive models. The primary difference is in the predictive information used. 18 © Deloitte Consulting, 2004 From Credit Scores to Predictive Models Credit scores are PMs that use only creditrelated variables to predict relative profitability. Payment pattern information, bankruptcies/liens, collections, inquiries, bad debt/defaults… But PMs can also be built using Both credit and non-credit information (preferred) Only non-credit information (perfectly feasible) 19 © Deloitte Consulting, 2004 Non-Credit PMs Why would we want to build a purely noncredit PM? Competitive advantages – e.g. matrix with credit State-specific regulatory constraints Expense of ordering credit reports Thin files/no-hits Public relations But from a purely actuarial POV, credit is predictive should be used as part of the PM! 20 © Deloitte Consulting, 2004 PMs: Considerations The key is to use as much information as possible in a multivariate way Choice of statistical techniques is important, but the real key is the quality and breadth of predictive variables used. GIGO Actuarial/insurance knowledge is critical Untapped riches reside in many companies’ transactional records. 21 © Deloitte Consulting, 2004 PMs: Data Sources We classify possible data sources into two groups Internal data sources: predictive information gleaned from the company’s own systems Regardless of how or whether it is currently used External data sources: predictive information available from 3rd parties. Both credit and non-credit 22 © Deloitte Consulting, 2004 Internal Data Sources Policy information Line-Specific information Limits, Deductibles, Measure of exposure (# cars, #houses, #employees, $sales, premium size… Driver, Vehicle, Business Class … Policyholder information Age, gender, marital status … 23 © Deloitte Consulting, 2004 Internal Data Sources Customer-level information Transactional data Coverage, premium and loss transactions Billing information Correlation with credit Agent information A little creativity in using these data sources will go a long way! 24 © Deloitte Consulting, 2004 External Data Sources Credit Predictive both for commercial and personal lines MVR – CLUE Zipcode/geographic information Rating territory Many different sources available The sky is the limit but Consider cost, hit rate, implementation, …etc 25 © Deloitte Consulting, 2004 Types of Variables Generated Territory-level Policy / policyholder-specific Demographic, weather, crime, ...etc Many traditional rating variables fall into this category Behavioral Less traditional – fits more neatly into data mining paradigm than classification ratemaking Credit, billing, prior claims, cancel-reinstatements… 26 © Deloitte Consulting, 2004 How Many Variables? It is possible to generate literally hundreds of predictive variables Some will be redundant Some will not be very predictive Some will be somewhat predictive Some will be “killer” A good model can contain as few as 15-20 or as many as 60-70 variables Usually no single “ideal” model 27 © Deloitte Consulting, 2004 Which Variables to Use? Choosing is a major part of the data mining process Use variety of exploratory statistical techniques Use prior modeling experience / actuarial knowledge Several considerations Actuarial / underwriting knowledge Client’s business needs Legal / regulatory considerations Data availability / cost Systems implementation considerations 28 © Deloitte Consulting, 2004 In Our Experience…. Do non-credit PMs work? YES: non-credit predictive models are Valuable alternative to credit scores Flexible Tailored to individual companies Leverage company’s untapped internal data Comparable predictive power to credit scores And mixed credit / non-credit PMs can be even stronger 29 © Deloitte Consulting, 2004 …But It’s Not a Walk Through the Park Challenges for PMs: IT resources constraints Project management Business process buy-in Success of system and business implementation Training and organizational change 30 © Deloitte Consulting, 2004 Conclusions © Deloitte Consulting, 2004 Industry Trends How do companies try to stay competitive regarding the use of credit? How do companies prepare for increasing regulatory constraints? Industry trends Companies are developing modeling capabilities and pursuing various applications Companies are developing proprietary credit scoring models rather than buying “off-the-shelf” credit scores. Companies are also going beyond credit, to build scoring models that don’t rely solely on credit 32 © Deloitte Consulting, 2004 Keys to Building PMs Fully utilize all sources of information Leverage company’s internal data sources Enriched with other external data sources Use large amount of data Employ systematic analytical process Use state-of-the-art modeling tools Apply multivariate methodology Disciplined project management 33 © Deloitte Consulting, 2004 Implications for Actuarial and Ratemaking Practice Opportunities for out-of-the-box thinking (who thought of credit a decade ago?) Increased multivariate analytic projects in the future On-going search for new predictive data sources, new modeling techniques, and new applications Next generation of pricing – more segmentation LTV, fraud, cross sell, retention, ..etc. A price for every risk New methodologies Statistical computing Lift curve concept Blind test / model validation methodology ROI benefit calculation …etc 34 © Deloitte Consulting, 2004 Implications for Ratemaking Principles Actuarial Ratemaking Principle #1: “A rate is an estimate of the expected value of future costs” Actuarial Ratemaking Principle #4: ” A rate is reasonable and Not excessive Not inadequate Not unfairly discriminatory But is that really the way profit-seeking companies price their products? Are rates ultimately based on costs or on what the market will bear? 35 © Deloitte Consulting, 2004 Implications for Ratemaking Principles How do you measure the ROI of a traditional ratemaking/class plan exercise? Why do ratemaking principles not mention blind tests of pricing algorithms? “Unfairly discriminatory”: If we develop a powerful new segmentation model, is it discriminatory to certain risks? If we don’t introduce it, is it discriminatory to other risks? How do we know if we don’t do the analysis? 36