Metodologie per Sistemi Intelligenti Data Mining Applications in the Italian Insurance Market Ing. Igor Rossini Laurea in Ingegneria Informatica Politecnico di Milano Polo Regionale di Como Agenda • Reference scenario and strategic framework • Cutomer Life Cycle • Data Mining Applications © Igor Rossini Strategic Framework Economic Scenario Global and International Market Integration (M&A) Processes Marketing Strategy Competitive Context Customer Segmentation Increased Level of Competition Italian Insurance Market New Client Needs and Behaviour New Channel of Distributions and Operators Reduction market growth rate Product Innovation Legal ITC Evolving Legislation Improvement of Technology © Igor Rossini Intensity of Competition Marketing Strategy Evolution Client Needs Market-Driven Management Selling Aptitude Product Concept Demand>Supply Demand=Supply Demand<Supply Maturity Level Market © Igor Rossini Customer Life Cycle Acquisition Activation Prospect Responder Relationship Management and Retention Client Ex Client High Value Target Market New Customer Customer Voluntary Churn High Potential Low Value Forced Churn © Igor Rossini Customer Life Cycle Events Prospect Responder Client Ex Client High Value Target Market New Customer Customer Voluntary Churn High Potential Low Value Forced Churn Acquisition Campaign Use Anti Attrition Campaign Response Acquisition Campaign Churn Info Requests Cross-Selling campaign Adhesion Up-Selling campaign © Igor Rossini Data Mining Application on Customer Life Cycle High Value Target Market New Customer Customer High Potential Low Value -Predictive model for Selling -Predictive model for Risk Analysis -Descriptive model on “Relevant” Attributes Voluntary Churn -Descriptive model on Customer Behaviour Forced Churn -Predictive model for Churn -Predictive Model for Cross/UpSelling campaign -Predictive model for Fraud Detection © Igor Rossini Swiss Life • Title: Innovative marketing strategies using Data Mining solutions • Challenge: support marketing initiative to preserve and extend the market share of the insurance company • Results: better prospect selection, an efficient churn analysis, new descriptive model for client segmentation © Igor Rossini Mining Environment ADLER © Igor Rossini ADLER carachteristic • Numerous data mining algorithms • User-friendly interface for data entry and for setting analysis criteria • MASY datawarehouse contains information on: – policies – social and demographic attributes – spending level of population for geographic areas © Igor Rossini Dr. Van Der Putten Case • Title: Data Mining in an insurance company • Challenge: expand the market for a caravan insurance product with low cost investment • Results: improvement in the selection of individual prospects and better description of existing customers © Igor Rossini Predictive Model • The model assigns to each customer a score meaning the purchase probability of the policy © Igor Rossini Descriptive Model • The model identifies, among all the customer of a caravan policy, interesting groups for marketing initiatives 3 4 2 5 1 © Igor Rossini Toro Assicurazioni • Title: Behavioral segmentation of retail customers • Challenge: characterize the purchasing profile of customers • Results: more efficient marketing initiatives, target product development, Life Time Value of customer knowledge © Igor Rossini Project Structure Data Base for Analisys Factor Analisys Clustering - Different clustering algorithm applied - Data Base development - Customer Table definition - Trade off cluster numerosity and his level of meaning - Factor analysis with no significative results Profiling Distance Map - Distance map of each client from the “centre” of his segment - Cluster description - Business Intelligence © Igor Rossini Customer segments of interest 1 AUTO FULL OPTION 2 AUTO BASIC 3 FUTURO E TUTELA 4 PENSO AI MIEI 5 AUTO E SALUTE 6 ALL BUSINESS 7 IN SALUTE 8 GIOVANI PREVIDENTI 9 POCHI MA BUONI 10 CASA E FAMIGLIA “N “DI TUTTO, DI PIU’” ” © Igor Rossini Farmers Insurance Group • Title: Driving profitability • Challenge: Data Mining to the insurance industry in large-scale profitability and risk analysis • Results: identification of “nuggets” of information “useful” for reducing frequency and severity of claims © Igor Rossini Project details • Data: 4 years of historical data, 2.4 million policies, 35 million records • Solution: Underwriting Profitability Analysis (UPA), a customer tailor package software developed for the insurance market by IBM, based on a decision tree algorithm © Igor Rossini Rules discovered • 40 “nuggets” of information useful to generate lower cost for claims of several million dollars • Example of a rule (illustrative) Rule #22 IF Field “VANTILCK” “Vehicle Antilock Break Discount?” = “Antilock Brake” Field “VEHTYPE” “Type of Vehicle” = “Truck” THEN claim rate 0,0115561 mean severity 5516,84 std dev severity 11619,9 pure premium 63,753 loss ratio 0,688204 608 training claims out of 53221 training points © Igor Rossini NRMA • Title: Insurance risk assessment using a KDD methodology • Challenge: Acquiring knowledge for the domain of motor vehicle insurance premium setting • Results: interesting pattern in the data useful to better insight policy premium setting © Igor Rossini Preprocessing © Igor Rossini Rules discovered • Example of a rule If And And And Then age < = 20 sex = male insured_amount > = 5000 insured_amount < = 10000 insurance_claimed = 1, cost = 0. (0, 15) • Claim associated with each risk area (fig. 1) and high claim risk area (fig. 2) Claims 1 2 3 4 5 … 16 Table 1 Number of rules 1090 494 192 82 38 … 2 Claims 15 16 22 Exposure 1745 2198 265 Cost 43308 50213 85678 Table 2 © Igor Rossini Australian Health Insurance Commission (HIC) • Title: Applying Data Mining Techniques to a Health Insurance Information System • Challenge: demonstrate the effectiveness of two data mining techniques in analyzing and retrieving unknown behavior patterns • Results: detection of patterns in the ordering of pathology services and classification of the general practitioners into groups reflecting the nature and style of their practices © Igor Rossini Neural segmentation © Igor Rossini Association Rule • the number of association rules obtained: Smin= 1% Cmin= 50% 24 rules Smin= 0.5% Smin= 0.25% 64 rules 135 rules • an example a rule: ‘If Iron Studies and Thyroid Function Tests occur together then there is an 87% chance of Full Blood Examination occurring as well. This rule was found in 0.55% of transactions.’ © Igor Rossini X- Insurance • Title: Data Mining techniques applied to motor auto policies • Challenge: better knowledge of customer claim profile to support marketing initiative for market share growth • Results: policy premium setting developed according to the level of risk of the customer group discovered © Igor Rossini Cluster discovered (1) CLUSTER # CUSTOMER % “Top Driver” 398.578 34,33% “Tradizionali” 275.742 23,74% “Donne In Carriera” 234.537 20,20% “Mix Alto Potenziale” 139.114 11,98% “Guidatori Inesperti” 113.080 9,74% © Igor Rossini Cluster discovered (2) © Igor Rossini “Guidatori inesperti” © Igor Rossini COIL CHALLENGE 2000 • Title: predicting and explaining Caravan Policy Ownership • Challenge: promote the application of computational intelligenge and learning technology to the real world problems • Task: – predict which customers are potentially interested in caravan insurance policy – describe the actual or potential customers; and possibly explain why these customers buy a caravan © Igor Rossini policy RESULTS • Prediction tasks: the winning model, based on a naive bayes approach, selected 121 policy owners on a total of 238 • Description Task: the winning model was built using the association rule method and better explained why people were not interested in a caravan policy © Igor Rossini Other Applications (1) INSURANCE COMPANY DATA MINING MODEL Predictive model for policy renewals 1- Predictive model to select the best customer for selling banking products 2- Predictive model for a cross-selling campaign Predictive model for churn analysis Descriptive and predictive model for policy rate setting © Igor Rossini Other Applications (2) INSURANCE COMPANY DATA MINING MODEL Descriptive model for behaviural customer segmentation Predictive model for fraud detection Predictive model for fraudulent medical services detection Descriptive model to discover pattern of interest among claims © Igor Rossini