Session 34 PD, Predictive Modeling Session 3 - Applications of Predictive Modeling. Moderator: Donna Christine Megregian, FSA, MAAA

Transcription

1 Session 34 PD, Predictive Modeling Session 3 - Applications of Predictive Modeling Moderator: Donna Christine Megregian, FSA, MAAA Presenters: Glenn Hofmann, MBA, Ph.D. Scott Anthony Rushing, FSA, MAAA Elliott Wallace

2 TransUnion / RGA Mortality Study Credit based solutions for Life Insurance Scott Rushing FSA, MAAA Vice President and Actuary, Global R&D RGA Reinsurance Glenn Hofmann Ph.D., MBA Senior Vice President, Analytic Services TransUnion SOA 2014 Life & Annuity Symposium Atlanta, GA May 19, 2014

3 Agenda Introduction / Motivation RGA Part 1 (Building the Mortality Models) TransUnion What is Credit Data? Insurance Adoption and Regulation Data / Modeling Methodology and Variable Examples Part 2 (Validating the Credit-based Mortality Models) RGA Analysis Methodology Results Summary Potential Applications 2

4 Introduction Introduction / Motivation TransUnion / RGA joint research to better understand the predictive nature of credit data for life insurance TransUnion built the Life Mortality Index (predictive model based on credit data) TransUnion shared the mortality index and about 85 other credit and demographic variables with RGA for a more traditional actuarial analysis Motivation for the Study Deeper understanding of credit data and the many credit scores available Desire to work with partners having data Insurance solutions for underinsured using data-driven analytics General R&D Predictive & Protective Value Studies to benefit the industry Rx MIB MVR Protective Term Tail Studies 3

5 TransUnion is a trusted partner for businesses and consumers around the globe Insurance Leading provider of credit and many other data types to the Insurance industry for over 20 years All of the top 10 P&C Insurance carriers are customers of TransUnion Full data/analytics product suite, e.g. Generic and Customized Insurance Scores, Vehicle History Score, A-Plus Risk Alerts, Risk Verification Platform (fraud solution) Comprehensive offering for Life Insurance Financial Services Insights and services to manage risk and grow profitability Eighteen of the top twenty U.S. banks and all major card issuers P&C Companies Healthcare Systems and solutions for the provider (hospitals) and payer (health insurers) markets Entered market in 2004 countries Consumer Other Industries Consumer credit reports and scores Identity monitoring and fraud prevention businesses Tenant screening services for property management firms and landlords Mortgage, auto, and employment screening Collections prioritization and recovery solutions for 15 of the top 20 collections companies 4

6 TransUnion in Life Insurance Underwriting Mortality Index (credit-based) Prescription data (and derivatives) (partnership) Violations data (exclusive partnership) Marketing Response and lapse models Acquisition and retention triggers Fraud and ID Identity verification/authentication suite Extensive fraud prevention solutions

7 Credit Reporting Process Consumer Collection Agencies Courts Lender/ Creditor 1 Lender/ Creditor 2.. Lender/ Creditor 10 Utilities Etc. TransUnion Comprehensive credit report on individuals (Scores, Attributes or Full file) Consumers Landlords Insurers (P&C, Life) Lenders Utilities Collection Agencies Employers (new hires) Public Records 6

8 What is on a Credit Report? Demographics Trades (Loans) Inquiries (hard) Public Records Name Address SSN DOB Credit cards Mortgages Auto loans Personal loans Consumerinitiated applications for credit Bankruptcies State/federal tax liens Civil judgments Phone Home equity loans Other debt At various points in time Balances Payments Limits Opening date Closing date Type Types Lender Name Dates Dates What is NOT on a Credit Report? Specific credit card purchase transactions IRS data Income, race etc. Checking/savings account data 7

9 Federal Statute Credit Regulations and Insurance Adoption Fair Credit Reporting Act (FCRA) Section 604 specifies permissible purposes for use of consumer reports to a person which it has reason to believe intends to use the information in connection with the underwriting of insurance involving the consumer Thus, consumer reports may be available to be used in connection with the underwriting of life insurance underwriting. State Laws and Regulation NCOIL Model Act ~ 50% of states adopt Model Act for P&C: Sensible restrictions on credit variables Most other states propose own rules for credit use in rating/underwriting 3 states (CA, MA, HI) do not allow credit P&C Insurers 99% of insurers use credit in connection with underwriting Life Insurers Early adopters are using in connection with underwriting Used in direct marketing by several life insurers 8

10 Credit Report Variables Scores Inquiries New trades Recency and Frequency Seeking Credit Credit Report The attributes encompass the breadth of the data Severity Recency Count Amount Negative Behavior: Public Records, Collections, Delinquencies, Past Due Tenure Most emphasis on longrun behavior rather than short-term patterns. Activity File thickness (e.g. number of active trades) Longevity (e.g. months since oldest trade) Utilization percentage Recency of use Usage patterns

11 Modeling Methodology - Modeling Process Data: Sample depersonalized dataset from the credit-active population in 1998 (90% of adult U.S. population). Non-overlapping samples from 4 quarterly credit archives (3/26/98, 6/26/98, 9/26/98, 12/26/98) to avoid seasonality. Exclusions: People with Age <20 and >=70, Missing/invalid SSNs, People w/o credit history Mortality Definition: Policyholders who died during the 12-year observation period (98-10). Merged deaths from Social Security Master Death File from October, 2011 using SSN Most credit features Variable Selection Modeling TransUnion Mortality Index Started with ~800 variables, all indicating features of credit history Variables with highest predictive power for mortality Low correlation among variables Stable variables (over time) Non-gameable variables Binary Logistic Regression Age at Issue, Implied Gender and Region used as control variables Score card development in multivariate context 0 (low risk) to 100 (high risk) Single number to express mortality risk on continuous scale Used for subsequent AE analyses 10

12 Credit Variable Examples: Number of Mortgages Mortality AE 130% 120% 110% 100% 90% 80% 70% 60% % of population Mortality AE Loss Ratio Rel from TU Auto Insurance Score Model 60.0% 50.0% 40.0% 30.0% 20.0% 10.0% % of population 50% Low High # Satisfactory Mortgages (active and positive payment behavior) E - Expected Mortality based on the model predictions for Age, Gender and State Higher # Satisfactory Mortgage trades is indicative of lower mortality ratio. # Mortgages is not as predictive in the TransUnion Auto Insurance Score model %

13 Credit Variable Examples: Number of Inquiries 150% 140% 130% 120% 110% 100% % of population Mortality AE Loss Ratio Rel from TU Auto Insurance Score Model 40.0% 35.0% 30.0% 25.0% 20.0% 15.0% 10.0% % of population 90% 5.0% 80% Low High 0.0% # Consumer-generated Inquiries in the past 2 years Moderate # Total Inquiries is indicative of lower mortality ratio compared to very low activity and high activity. Increasing trend for the same variable in the TransUnion Auto Insurance Score model illustrates that TransUnion Mortality Index is very different from other TransUnion Scores. 12

14 Data Flow Sampled from 1998 credit active population Death rate during the 12 yr observation period (98-10) ~ 6.5% 92 M Variables in the file Age, gender, state, date of birth, death date, credit variables Model Build & Testing Random Holdout Train 44 M Validation 30 M 18 M Used by TransUnion for Model building and testing. Results are credible because of the big sample size Final Result TransUnion Mortality Index 13 RGA used the TransUnion Mortality Index for the Mortality Study

15 Study Background Analysis Methodology Studied a random sample of nearly 18 million individuals from the 1998 credit archives. TransUnion Mortality Index was calculated for each individual. RGA built a 12-year mortality study ( ) on the 18 million lives. Study resulted in 194 million exposure years and 1.1 million deaths (using the Social Security Master Death File from October, 2011) Gender Variable: Derived based on the individual s name (one gender required > 90% of the time). Those names with < 90% (unknown gender) were excluded. Expected Basis: Started with historical US Population mortality tables (by gender & attained age, includes improvement) Adjusted the Expected for: Gender mixing in the data Missing SSMDF deaths by age and calendar year 14

16 Overall Results Exposure distribution is nearly uniform across the index Smooth increasing AE curve Results 300% Overall Mortality 20 AE (Adjusted Population) 250% 200% 150% 100% 50% 0% Exposure Count in Millions TU Mortality Index 15

17 Results Results by Issue Age Similar shapes by age (in 1998), but slope seems to flatten for older ages Increasing mortality (as % of pop mort) as age increases Mortality by Issue Age 250% AE (Adjusted Population) 200% 150% 100% 50% % TU Mortality Index 16

18 Results Results by Gender Adjustments to Expected are not perfect by gender, so graphs are displayed separately. Mortality comparison (M vs. F) not appropriate, but shapes are similar Mortality for Males Mortality for Females 250% 250% AE (Adjusted Population) 200% 150% 100% 50% AE (Adjusted Population) 200% 150% 100% 50% 0% % TU Mortality Index TU Mortality Index 17

19 Results Results by Duration Appears to be a similar pattern for all durations in the study AE (Adjusted Population) 250% 200% 150% 100% 50% 0% Mortality by Duration TU Mortality Index

20 Results Results by Duration Splitting policy years into early duration versus later duration, the slope does flatten a little... but not much. Mortality by Duration 250% AE (Adjusted Population) 200% 150% 100% 50% % TU Mortality Index 19

21 Results Results by Region Similar shapes by region. Region is not directly used in the model, but credit data and behaviors do vary by region. This is reflected in the Average Mort Index. AE (Adjusted Population) 250% 200% 150% 100% 50% Mortality by Region Midwest Northeast South West Region A/E Average Mort Index Midwest 97% 50.0 Northeast 87% 46.3 South 106% 54.1 West 91% % TU Mortality Index

22 Vantage 3.0 Score (traditional credit score) vs. TransUnion s Mortality Index Rows/ columns represent 2% of the records in that score Dots are centered on the mean (x and y) of each group Size of circle represents total exposure in the group A lot more variance in the TransUnion Mortality Index for the higher Vantage scores Vantage Score Results TransUnion Mortality Index

23 < Results Results split by Vantage Score Band Individuals are grouped into 3 categories according to Vantage Score TransUnion Mortality Index separates similar credit risks by AE (Adjusted Population) mortality risk 250% 200% 150% 100% 50% 0% Mortality by Vantage Score TU Mortality Index Exposure in Millions

24 Potential Applications Target Marketing Can be used to assist in the market segmentation process Among other things, focus on the best risks for new customers or upsell / cross-sell campaigns Conversion near the end of the level-term period Assist in the selection of policies for conversion near end of term Offer favorable conversion terms to less risky policy holders Simplified issue programs Use in conjunction with other real-time data (violations, Rx, MIB) Additional segmentation in full underwriting Lapse prediction and related underwriting actions (premium, face, payment terms, etc.)

25 Summary Summary / Next Steps TransUnion Built a Life Mortality Index based on credit data this is not the same as traditional credit score (Vantage score or FICO score) Exploratory research conducted by RGA to understand the data and models RGA s validation using a 12-year traditional actuarial mortality study Study Results Score appears to be very predictive of mortality across many dimensions A/E of the worst decile more than 4x higher than that of best decile Evidence does not appear to wear off very fast Many potential applications in mind Next Steps Complete the paper (next month or so) Test the model index results on insured lives

26 Scott Rushing FSA, MAAA RGA, VP Global R&D Glenn Hofmann Ph.D., MBA TransUnion, SVP Analytics Questions 25