Predictive Modelling Not just an underwriting tool Joan Coverson 15 May 2013 Predictive Modelling - Agenda Predictive Modelling High level overview Some uses: Lapse propensity Geodemographics Business mixes 2 1
Predictive Modelling How does it work? Take data where you know results Build a model to fit Use model to predict unknown results 3 Predictive Modelling in General Insurance Personal Driving history Type of car Age Gender Marital status Address etc Credit score Length credit history Payment history Debts etc Internal data External data 4 2
14/05/2013 Retention Modelling 5 Retention Strategies Passive Reactive Proactive Predictive Modelling 3
Retention Case Study UK bancassurer Large block of life and CI business Lapses modelled Understand drivers Target likely lapsers Lots of lapses Good for predictive modelling Approx 200,000 policies 244 different fields of data Policy + Banking 7 Some General Principles Try and identify main variables & eliminate what you don t need Produce model using single variables, eliminate non-significant Produce model with interactions Check fit and refine variables if necessary Produce model, refine etc 8 4
Step 1 Try and identify main variables & eliminate what you don t need First name Produce model using single variables, eliminate non-significant Produce model with interactions Check fit and refine variables if necessary Produce model, refine etc 9 Step 2 Try and identify main variables & eliminate what you don t need Produce model using single variables, eliminate non-significant Produce model with interactions Check fit and refine variables if necessary Produce model, refine etc 10 5
Example - Age band (Significant) Parameter estimate age band Parameter estimate 0.6 05 0.5 0.4 0.3 0.2 0.1 0 0.1 0.2 0.3 0.4 0.5 A:<=30 B:<=40 C:<=50 D:<=60 E:<=65 F:<=70 G:>70 700,000 600,000 500,000 400,000 300,000 200,000 100,000 0 Record count 11 Example Credit Card Holdings (not significant) Parameter estimate credit cards held 0 0.50 1 Parameter estimate 1.5 2 2.5 3 3.5 4 0 1 2 3 4 Credit Cards 1,200,000 1,000,000 800,000 600,000 400,000 200,000 0 12 6
Step 3 Try and identify main variables & eliminate what you don t need Produce model using single variables, eliminate non-significant Produce model with interactions Check fit and refine variables if necessary Produce model, refine etc 13 Interactions Lapse Example Lapse Rate Males 13% Females 12% Lapse Rate Non-Smokers 12.5% Smokers 12.5% Lapse Rate Male non-smokers 15% Male smokers 5% Female non-smokers 10% Female smokers 20% 14 7
Step 4 Try and identify main variables & eliminate what you don t need Produce model using single variables, eliminate non-significant Produce model with interactions (NB need example) Check fit and refine variables if necessary Produce model, refine etc 15 Example - Occupation code 1.5 1 Parameter estimate - occupation code? What about combining codes 700000 600000 Parameter estimate 0.5 0-0.5-1 500000 400000 300000 200000 100000-1.5 0 2 4 6 8 10 12 14 16 18 20 23 25 27 30 0 16 8
Step 5 Try and identify main variables & eliminate what you don t need Produce model using single variables, eliminate non-significant Produce model with interactions (NB need example) Check fit and refine variables if necessary Produce model, refine, produce model, refine 17 Annual Predicted Lapse Rates (sample) 1800 1600 1400 Frequency 1200 1000 800 600 Frequency 400 200 0 Lapse Rates 18 9
South African Case Study Large SA protection writer Whole of Life Mortality business 800,000 policies Modelling similar to UK example but Split data up into monthly chunks Added external economic data 19 Predicted vs. Actual 4.00% 800 3.50% 700 3.00% 600 Actual 2.50% 2.00% 1.50% 1.00% 500 400 300 200 Actual Theoretical 95% CI 95% CI Lapses 0.50% 100 0.00% 0 0.00% 0.50% 1.00% 1.50% 2.00% 2.50% 3.00% Predicted 20 10
Predicted vs. Actual Lapses Predicted Actual 2002/01/012003/01/012004/01/012004/12/312006/01/012007/01/012008/01/012008/12/312010/01/012011/01/012012/01/01 21 Inflation & GDP 15% 180% 160% 10% 140% Inflation & GDP 5% 0% -5% Peak in expected lapses when inflation at peak and GDP at its lowest 120% 100% 80% 60% 40% 20% Lapse Effect Inflation3 GDPChange6 Lapse Effect -10% 0% 2012/04/01 2010/11/18 2009/07/06 2008/02/22 2006/10/10 2005/05/28 2004/01/14 2002/09/01 2001/04/19 Month 22 11
14/05/2013 Retention Modelling Other information about the life assured Traditional factors Income Household status Credit scores Health... Term Duration Smoking Distribution... External information Economic indices Stock market indices Predictive Model Geodemographic Mortality 24 12
Predictive Model Policyholder addresses Postcodes Demographics of postcode Mortality / morbidity for specified postcode 25 UK Geography UK Country Lower Super Output Area (E&W) Data zone (Scotland) Something Else (NI) Census Output Area Postcode 13
Postcodes 2.5m+ Approx 15 addresses Links to Mosaic (Experian) Acorn (CACI) Census Output Areas On average 11 postcodes England 214 adults (20-4000) Scotland 88 adults (5-2350) Census Data Education Socio Economic Class Home type Etc. 14
Lower Super Output Areas OA Grouped Together Automated Set minimum size = 1000 Average 1500 Data Zone 500-1000 Key small area units Links to Index for Multiple Deprivation http://neighbourhood.statistics.gov.uk Index of Multiple Deprivation 4 Available: England, Wales, Scotland & NI Lower Super Output Area (England) Measures deprivation Domains Income Employment Health Education, Skills & Training Access to Services Housing Crime Different weights and actual measures in each country 15
Census Uses output areas (lower level) Map postcodes to output t areas Multiple variables Predictive model Identifies significant and non-significant variables Models mortality outcomes (or underwriting assessments) Needs lots of data! 31 Results Mortality Example Level of highest education proved an important variable 50% variation after allowing for expected differences by SA Comparable to gender variation Greater than SA variation 32 16
14/05/2013 Applications of Postcode Modelling Pricing Reporting R ti Marketing Target areas for marketing Special offers Distribution Understand characteristics of business by distributor 33 Business Mix Modelling 34 17
Predictive Modelling Gender Mixes Initial model produced pre 21/12/2012 One way splits only provide limitedit information Predictive model considers multiple interactions e.g. Sum assured and age Age and term Output = % males for combination of multiple variables 35 Significant and Non-significant variables Significant variables Occupation class Age Cease age Policy term Family status Sum assured Product type (e.g. LTA, DTA) Smoker status Insignificant variables Policy year Policy month 36 18
Final Variables and Combinations Used Variables in Final Model 1) Cease age + age at entry 11) Product type + term 2) Cease age + term 12) Product type + sum assured 3) Family status + age at entry 13) Term 4) Family status + term 14) Term + age at entry 5) Family status + sum assured 15) Sum assured 6) Age at entry 16) Sum assured + term 7) Product type 17) Smoker status + cease age 8) Product type + cease age 18) Smoker status + family status 9) Product type + family status 19) Smoker status + product type 10 ) Product type + age at entry 37 Model Fit 1 0.9 0.8 0.7 0.6 0.5 0.4 Actual male % Expected male % LCL HCL 0.3 03 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 38 19
Model Fit excl Family Status 1 0.9 0.8 0.7 0.6 0.5 0.4 Actual male % Expected male % LCL HCL 03 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 39 Sample Output Rating factor Value Age Next Birthday at entry (whole years) 45 Policy Term (whole years) 20 Smoker status Y Sum assured 150,000 Cover type Level Term Assurance % Male 56% 40 20
Variations Rating factor Value Age Next Birthday at entry (whole years) 45 Policy Term (whole years) 20 Smoker status Y Sum assured 500,000 Cover type Level Term Assurance % Male 71% 41 Uses for Business Mix Model Pre G-day Predicting mix for pricing Now Monitoring mixes Targeting specific segments Predictive modelling can be used to predict any type of mix (not just gender) 42 21
Conclusions Predictive Modelling enables you to get more out of your data You can use it for almost anything! All you need are a few clever actuaries! 43 22