Harnessing Predictive Analytics to Improve Customer Data Analysis and Reduce Fraud Patrick Shearman General Manager, Information Management HCF of Australia Ltd Technology and Innovation for Insurance March 2007
Introduction What is Predictive Analytics? Targeting individuals customers more effectively Streamlining, simplifying and accelerating data analysis An overview of some techniques used in fraud detection Strengths and weakness Conclusion
What is Predictive Analytics? Originated in statistical related research studies Predictive Analytics Advanced Analytics Statistics Data Mining Text Mining Data Visualisation Reporting Decision Optimisation Rules Engine Scoring Engine Looking for patterns of behaviour within historical data BI analysts use the output in conjunction with traditional analytical techniques
Enterprise Platform for Predictive Analytics Enterprise Data Sources Marketing Attitudinal Interaction Web Call-center Operational Unified Data View Understand Predict Reporting Statistics Data mining Web mining Text mining Advanced Analytics Analytical Asset Management Act Campaign optimization Interaction optimization Real-time / Batch scoring Real-time risk assessment Feedback management Decision Optimization Multi-channel CRM Acquisition Cross-sell Customer Service Predictive routing Service quality Market, Product, and Promotion Planning Brand perception Basket analysis Product design Customer feedback & satisfaction Finance and Risk Fraud/AML/Credit Shrink Legal compliance Financial/sales Actuarial/underwriting budget variance Operations Supply chain analysis Category/inventory management Demand forecasting Retention Segmentation Promotion & marketing mix Site location Medical outcomes Human resources Customer Contact Channels Website Email Phone Mail Branch ATM Agent Mobile
Customer Segmentation Model Value / Risk Behavioural Categorisation HCF CUSTOMER BASE Predictive Analytics Integrated and ongoing Identify Customer Risk Dimensions SALES AND Identify MARKETING Customer Value Dimensions START UP AND PERIODIC Analyse Variables Determine Rules Categorise Customers by VRC PBI Develop and Implement VRC Model Monitor Campaigns And Customer VRC SALES AND MARKETING Profile VRCs And append Customer research ONGOING Auto Update Customer VRC Develop and Implement Action Plans Execute Campaign Generate Campaigns And Targets PBI
Customer Segmentation Model Value / Risk Behavioural Categorisation Predictive Analytics Sales and Marketing Program Examples Targeted Acquisition Programs HIGH POTENTIAL VALUE Life-Stage/Life-Event Programs TO ACQUIRE QUALITY NEW CUSTOMERS THAT MEET THE CURRENT HIGH VALUE / LOW RISK PROFILE TO BUILD LOYALTY IN THE HIGH AND MEDIUM RISK YOUNG SINGLES Retention Programs TO COUNTER THE LAPSE POTENTIAL OF ALL HIGH VALUE CUSTOMERS Upgrade Campaigns TO CAPTURE THE POTENTIAL OF ALL HIGH PROPENSITY MEDIUM VALUE CUSTOMERS Customer-Satisfaction Programs Guarantee LOW VRC 1 VRC 2 RISK VRC3 HIGH TO UNDERSTAND THE DRIVERS AND LEVELS OF CUSTOMER SATISFACTION IN HIGH RISK GROUPS New Customer Care Program TO LOWER ATTRITION RISK IN HIGH RISK ACQUISITION GROUPS
Modelling Techniques C5.0 example Models which are mathematical equations that describe the relationship among a set of data. Models have been built to predict member attrition, product up-sell and fraud. C5.0 algorithm also produces a set of rules that tries to make predictions for individual customers. C5.0 attaches a probability of a particular behaviour occurring ranging between 0 (will never occur) and 1 (will definitely occur). C5.0 models work well even in the presence of problems such as missing data and large numbers of input fields. In addition, C5.0 models tend to be easier to understand than some other model types, since the rules derived from the model have a relatively straightforward interpretation
Modelling Techniques C5.0 example Migration Analysis to analyse the natural product migrations of HCF members over the past two years. Predictive modelling to select product migration that provide high increases in premium and target members most likely to make those migrations in the near future. Years of transactional data Scoring & Strategies Predict Up-Sell 2 Years 3 Months 6 Months DATA VARIABLES Timing of upgrades Length of Membership Single vs. Family scale (children vs. empty nesters) Claims history for both hospital and extras Per Night Excess or co-payments Payment methods and frequency Highest level of Rebate entitlements Age of contributor (and dependents) Last product movement Life-time health cover loading Health Premium $ Demographic data
Rules for 1 - contains 18 rule(s) Predictive Analytics Modelling Techniques C5.0 example Rule 1 for 1 (51; 0.981) if LHC_LOADING_CHGE_FLAG = 1 then 1 Rule 2 for 1 (22; 0.958) Customers in Segment Probability of buying product if AGE <= 37 and NEW_SEGMENT = JC and PEEK_EXCESS_PROD = 0 and MODALITY_O_OOP_Count <= 1 and CLAIM_MONTHS > 4 then 1 Rule 3 for 1 (201; 0.946)
Modelling Techniques C5.0 example From the Gains Chart it can be seen that 60% of those customers most likely to respond will be achieved in the first 10% of customers mailed, compared to about 10% if data mining was not used. More important than the Gains Chart is the Response Chart which indicates a response of rate of 3% if we mailed the first 10% of customers. The response rate is much higher than the baseline response rate of about 0.5% which one would get without data mining.
Customer Segmentation Model Value / Risk Behavioural Categorisation HCF CUSTOMER BASE Predictive Analytics Integrated and ongoing Identify Customer Risk Dimensions SALES AND Identify MARKETING Customer Value Dimensions START UP AND PERIODIC Analyse Variables Determine Rules Categorise Customers by VRC PBI Develop and Implement VRC Model Monitor Campaigns And Customer VRC SALES AND MARKETING Profile VRCs And append Customer research ONGOING Auto Update Customer VRC Develop and Implement Action Plans Execute Campaign Generate Campaigns And Targets PBI
Predictive Marketing 1. Configure Predictive Analytics Predictive Marketing System Flow Clementine 3. Create predictive models 4. Deploy Model?! contact center Interaction Builder 9. Return customer data + SCORE 5. Incoming Call 7. Request Score Customer View Builder Clementine Server PEV Predictive Enterprise View 2. Read Data Real-time Server 8. Return Score Call Center Application 6. Read Customer data (7a) Read additional data DWH Lists Other, legacy, etc.. Transactional Operational Predictive Enterprise Services PER Analytical Views (PEV) Scenario s Predictive Models Reports
Leverage inbound interactions Enterprise Data Sources 2 Real Time Marketing what is the best offer for this member? 1 Quick Survey to understand member s needs and preferences Marketing Attitudinal Interaction Web Offers A B Business rules Predicted Behavior 74% Value 90 C Call-center Operational C 32% 200 3 Dynamic Scripts to assist agents in making offer to the member
Targeting individuals customers more effectively Summary Can lead to a better understanding of your customer needs Integrated environment to drive interactions across all channels (inbound and outbound) Based on preferences, needs and risks of individual customers (Advanced Analytics) Determine the optimal contact mix of product, channel and time (Decision Optimization) Leverage of existing BI architecture Develops business rules around customer segments
Enterprise Platform for Predictive Analytics Enterprise Data Sources Marketing Attitudinal Interaction Web Call-center Operational Unified Data View Understand Predict Reporting Statistics Data mining Web mining Text mining Advanced Analytics Analytical Asset Management Act Campaign optimization Interaction optimization Real-time / Batch scoring Real-time risk assessment Feedback management Decision Optimization Multi-channel CRM Acquisition Cross-sell Customer Service Predictive routing Service quality Market, Product, and Promotion Planning Brand perception Basket analysis Product design Customer feedback & satisfaction Finance and Risk Fraud/AML/Credit Shrink Legal compliance Financial/sales Actuarial/underwriting budget variance Operations Supply chain analysis Category/inventory management Demand forecasting Retention Segmentation Promotion & marketing mix Site location Medical outcomes Human resources Customer Contact Channels Website Email Phone Mail Branch ATM Agent Mobile
Fraud Detection Analyze claims payment patterns of behaviour Customer Provider Staff Detection of inappropriate practice Using automation and data mining techniques Develop business rules to automate a monitoring process
Modelling Techniques for fraud detection Is the classification of objects into different subsets of data (clusters), so that the data in each subset (ideally) share some common trait Models which are mathematical equations that describe the relationship among a set of data. Highlights anomalies in the patterns of behaviour Data visualization C5.0 algorithm also produces a set of rules that tries to make predictions for individual clients. C5.0 attaches a probability of a particular behaviour occurring ranging between 0 (will never occur) and 1 (will definitely occur). C5.0 models work well even with missing data and large numbers of input fields, minimal model training times and easier to understand than some other model types, with rules relatively straightforward to interpret.
Fraud Model Clustering 3,121,842 Cases for 2 years data and 11 clusters
Fraud Model Anomaly 209,149 records in the cluster with 2,532 Anomalies (1.2%) Variables
Fraud Model Data visualisation Distinct Clustering Indistinct Clustering Anomalous Clustering Anomaly
Fraud Predictive Analytics Summary Automates the fraud detection process Provides profiles of customer and provider claiming patterns Detects anomalies or patterns of behaviour differing from the norm Identifies staff, customer and provider risks Helps build business rules around inappropriate practices
Conclusion - Strengths Simplifies the actuarial analysis process Uses proven statistical methodology Adds value to the overall business intelligence framework Gives better understanding of customers and helps target the right customers Challenges peoples current perceptions about a subject Develops business rules and models which can be integrated into systems architecture
Conclusion - Weakness Expensive Models can be high maintenance Can get complex Data preparation can be time consuming Results abstract and difficult to interpret May not discover anything new Return on investment can be marginal
Thank you & Questions?