Predictive Analytics:
Achieving Greater Decision Accuracy, Better
Risk Segmentation, and Greater Profitability
Lamont D. Boyd, CPCU, AIM
Insurance Market Director FICO Scoring Solutions
Agenda
ยป Overview/Research Findings
ยป 128 patents โ predictive analytics and decision
management
ยป Insurance Scores
ยป Credit-Based Insurance Scores
ยป Property Risk Scores
ยป Safe Driving Scores
ยป Custom Predictive Analytics/Custom Modeling
ยป Risk Segmentation for Marketing, Underwriting,
Product Development, Pricing, Claimsโฆ..
3
Predictive Analytics: Achieving Greater Decision
Accuracy, Greater Profitability
Towers Watson Survey (4Q/2011)
ยป Percentage of insurers currently using or planning to use
predictive modeling for underwriting and/or pricing
ยป 88% - Personal Auto ยป 83% - Homeowners ยป 73% - Commercial Property/CMP/BOP ยป 72% - Workersโ Comp ยป 63% - Commercial Auto ยป 52% - General Liability
ยป Why the growth in predictive modeling?
ยป 85% - Rate accuracy
ยป 69% - Loss ratio improvement ยป 69% - Profitability
Predictive Analytics: Achieving Greater Decision
Accuracy, Greater Profitability
SMA (Strategy Meets Action) Research (2012)
ยป 75% of P&C insurers will increase spending on analytics over the
next three years
ยป 60% currently using and 25% implementing analytics in underwriting to better understand and manage risks
ยป 50% currently using and 22% implementing analytics for distribution management/agent performance
ยป 40% using and 15% implementing analytics for claim fraud detection and prevention
ยป Personal lines carriers/units more aggressively using and pursuing predictive analytics than commercial lines counterparts
ยป Noted barriers to capitalizing on predictive analytics
ยป Lack of strategy
Agenda
ยป Insurance scores: CBIS and property risk scores
ยป Custom models: GAM versus GLM
ยป Premium leakage and pre-claims fraudulent
detection at underwriting
detection activity at
underwriting
Credit-Based Insurance Scores
ยป Key industry success!
ยป FICO introduced Credit-Based Insurance Scores (CBIS) to the
industry in 1993
ยป CBIS scores currently used by ~95% of all auto and home
insurers
ยป FICO CBIS scores are monitored for effectiveness re risk
segmentation and amended as required by regulatory/legislative
mandates
ยป Current generation FICO CBIS scores
ยป Experian/Fair Isaac Insurance Score 2.0 (via Experian/LexisNexis)
ยป InScoreยฎ 3.0 (via Equifax)
ยป Fair Isaacยฎ Insurance Risk Score 2.0 (via TransUnion)
Update:
FICO CBIS
in Current Economic Environment
Key Question:
Have FICO CBIS risk segmentation performance/score
distributions changed over time?
ยป Ongoing FICO and client analysis reveals continuing high levels
of performance
ยป FICO CBIS score distributions remain essentially the same
ยป Some consumers are seeing lowering scores
ยป Economic climate-driven factors
ยป Majority of consumers are seeing slightly rising scores
ยป More cautious credit management practices
ยป No late payments
ยป Opening fewer accounts
CBIS Usage
ยป CBIS usage provides consistent link in all areas that require
strategic decisions
Decision Engine Command Center
Credit-Based Insurance Score
Underwriter Management
Field Office Management Production Source Management
Customer Acquisition
New Business Underwriting Tier Placement
Renewal Underwriting External Info Purchase
Pricing
Cross-Selling
CRM
Billing / Reinstatement Fraud Detection
Acquisition/Marketing โ Client Benefits
ยป New Customer Acquisition/Marketing Strategy
ยป Focused application
ยป Cost-effective acquisition strategies
ยป Targeting most likely responders
ยป Response modeling
ยป Seeking greatest profit potential
ยป Focused/limited utilization currently ยป Few regulatory restrictions
Underwriting and Pricing โ Client Benefits
ยป New Business Underwriting Strategy
ยป Earliest CBIS application
ยป Consistent and objective new business risk segmentation decisions
ยป Quickly accepting applicants with greatest profit potential
ยป Dedicating resources to appropriate risks
ยป Minimizing exposure to riskier applicants
ยป Regulatory considerations
ยป Pricing Strategy
ยป Currently focused CBIS application
ยป Slotting applicants by price/tier based on multiple risk characteristics
ยป Gaining premiums relative to presented exposures
ยป Consistent and objective pricing decisions
New Business Underwriting Strategy
Decline Referral Automated Approval
1 2 3 4 40 50 60 70 80 90 100 110 120 130 140 0 5 6 7 8 9 10 L o ss Rat io Population by deciles Average
Renewal Underwriting โ Client Benefits
ยป Renewal Underwriting Strategy
ยป Secondary CBIS risk segmentation application
ยป Updating CBIS scores at renewal where allowed
ยป Identifying renewal policyholders for swift processing
ยป Highlighting renewal policyholders requiring greater attention
ยป Renewal tier-placement and pricing relative to exposure
ยป Consistent and objective renewal underwriting decisions
CBIS Enhanced Uses โ Client Benefits
ยป CBIS enhances decision-making and profitability for the
enterprise
ยป Book management
ยป Managing books of business at varying level for profitability
ยป Underwriting management
ยป Managing underwriter responsibilities to match risk complexity and varying levels of production and profitability goals
ยป Distribution source management
ยป Managing at agency or other production source levels to assure a focus on greater profitability through better managed and strengthened
Regulatory Environment Forecast
ยป Regulatory/legislative questions will continue
ยป Individual state actions
ยป NAIC committees/task forces
ยป FTC StudyโAuto (2007) and Home (2012/2013)
ยป Education is the key
ยป Education to help consumers understand and change habits to influence credit-based insurance scores is available at
www.insurancescores.fico.com
ยป Education to help consumers understand and improve their credit habits to influence the FICOยฎ scores lenders use is available at
www.myfico.com
ยป Under the 2003 Fair and Accurate Credit Transactions Act (FACT Act), consumers can access each credit report annually via
Property Risk Scores
ยป Property PredictRยฎ Insurance Scores โ launched in 2007 to help
insurers harness predictive value from property inspections
performed by FICOโs partner โ Millennium Information Services
ยป Snapshot of risk based on property inspection data ยป Rank-orders properties by loss ratio relativity
ยป Precise and objective measurement of a property risk
ยป Property PredictR provides strong additive value to CBIS or
where regulatory constraints make credit difficult to use
ยป Property PredictR scores are used as a key risk predictor in
conjunction with other risk variables
1.26 1.10 0.86 0.76 0.64 0.92 1.53 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 Below 180 11.7% 180-201 9.1% 202-222 19.3% 223-244 17.6% 245-272 20.7% 273-293 11.1% 294 or more 10.5% Score by Percent Premium
Los s R a tio R e la tiv it y
Property PredictR
ยฎScore Retro Analysis Loss Ratio
CBIS with Property PredictR
ยฎLoss Ratio
Relativity Matrix
SCORE RANGE
0.56 Consistently Good Risks
Highest Scoring 20% 1.37 0.71 1.18 1.11 1.83 1.85
Missing Credit Score
1.00 0.70 0.87 0.93 1.10 1.41 Total 0.47 0.66 0.66 Highest 20% 0.77 0.58 0.59 0.73 1.52 Medium High 20% 0.89 0.68 0.86 0.63 1.65 Medium 20% 0.97 0.65 0.73 0.97 1.83 Medium Low 20% Medium High 20% Medium 20% 1.48 Consistently Poor Risks
Lowest Scoring 20% Total Medium Low 20% Lowest 20% SCORE RANGE SCORE RANGE 0.56 Consistently Good Risks
Highest Scoring 20% 1.37 0.71 1.18 1.11 1.83 1.85
Missing Credit Score
1.00 0.70 0.87 0.93 1.10 1.41 Total 0.47 0.66 0.66 Highest 20% 0.77 0.58 0.59 0.73 1.52 Medium High 20% 0.89 0.68 0.86 0.63 1.65 Medium 20% 0.97 0.65 0.73 0.97 1.83 Medium Low 20% Medium High 20% Medium 20% 1.48 Consistently Poor Risks
Lowest Scoring 20% Total Medium Low 20% Lowest 20% SCORE RANGE Cr ed it -Based In su ran ce S co re
Note - Statistic listed in cell is the loss ratio relativity for that cell
Poor Risks Medium Risks Good Risks
Safe Driving Scores โ UBI/Telematics
ยป FICO Safe Driving Scores available in partnership with
DriveFactor, a leading telematics provider
ยป Continuous view of risk based on driving behavior ยป More closely align pricing to loss potential
ยป Accurate measurement of an automobile risk
ยป Consumer-facing views to gradually improve driving habits
Agenda
ยป Insurance scores: CBIS and property risk scores
ยป Custom models: GAM versus GLM
ยป Premium leakage and pre-claims fraudulent
detection at underwriting
detection activity at
underwriting
Custom Predictive Analytics/
Custom Modeling
Insurance Analytic Solutions
Insurance Industry P&CPersonal Health Homeowners P&C Commercial Auto Government Federal Medicare State Medicaid Fraud Detection Loss Reserving SubrogationRisk Scores Rules Mgmt.
Rules Mgmt, UW Profit, Claims
Rules Mgmt. Fraud Detection
Property & Casualty
Bureau Scores Risk Scores UW Decisioning Bureau Scores Risk Scores UW Decisioning
Property & Liability Workersโ Comp
Private Payors Life
Development &
Implementation
Model Development Methodology
Performance Definitions
Project Design
Generate Complex Characteristics
Model Development & Testing
Delivery Meeting Data Analysis Sampling Plan Performance Def Options Analytic Review Implementation/Tracking Plan Timing Analysis Final Performance
Definitions Review & Selection
Sample Selection โPull & Audit Start Up
documentation Customer & FICO Project Team
Meeting Project Specifications
Characteristic Selection
Analytic Documentation
Segmentation Analysis
Analytics Spectrum
Different decisions lead to different methodologies
Benefit
โข Brings all predictive analytics into a single
decision framework โข Assigns the optimal
action for each prospect/account โข Creates micro segments by matrixing 2 or 3 predictive models โข Rank orders prospects on a single dimension โข Establishes broad segments based on customer profile data HIGH Multi-Dimensional Trade-Off Assessment Predictive Models or โScoresโ Profiling & Segmentation X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X X Decision Optimization Benefit
โข Brings all predictive analytics into a single
decision framework โข Assigns the optimal
action for each โข Creates micro segments by matrixing 2 or 3 predictive models โข Rank orders prospects on a focused dimension โข Establishes broad segments based on customer profile data
Custom Analytics/Modeling
A model that will improve decisions should be โฆ
ยป Flexible
ยป Can capture complex data relationships ยป Palatable
ยป Interpretable
ยป Satisfies operational constraints
ยปLegal concerns
ยปImplementation friendly
ยป Able to be computed quickly and seamlessly deployed ยป Robust into the future
Project design
Understanding business challenges
ยป Data Mining - โTry multiple algorithms and see which one best matches the business challengeโ
ยป Objectives - โUnderstand the
business problem and design an
Prepare Modelling Datasets Visualise & Explore Build Standard Models Forecast Predictive Performance Access Data Deploy Analytics in Decision Build Scorecard Models ๏ Data Mining ๏ ๏ Decision ๏
Model Development
Initial data analysis
Fine Binning
Coarse Binning
ยป What are we trying to predict?
ยป Metrics
ยป How do we know the model is working?
ยป What data sources are available?
ยป What are costs vs. benefits?
ยป Common data issues
ยป Incomplete and/or inaccurate data ยป Outlier anomalies can skew statistics
ยป Highly correlated items blur the incremental value contributed by each predictive variable ยป Is the data biased? Censored?
Model Development
Segmentation
Segment 1 Scorecard 1 Segment 1.1 Segment 2 Total Applicant PopulationScorecard 2
Segment 1.2 Segment 2.1 Scorecard 3 Scorecard 4 Segment 2.2 0 1 # Scorecards 3 4 5 .066 .064 .062 .060 .058 .056 .054 .052 Low est C lai m Fr equency 2 6 .050 .068 .080 ...
โฆ how many scorecards to deploy:
Which segments matter most? Where the ROI is justified?
Model Development
Score Engineering
ยป Imposing constraints on the score formula to achieve various objectives ยป Objectives
ยป Palatability
ยป Legal requirements
ยป Robustness over changing times
ยป Adjustments for known sample biases ยป Engineering techniques
ยป Choose the โrightโ objective function โ possibly more than one ยป Individual weight restrictions
ยป Characteristic restrictions
Modeling Repository
Z
Implementation
Can I make it work in practice?
xโ X Y Runtime Package (Java) Modeling Process
IT Model Deployment Process
Data Preparation Variables Scoring Stats Data Mapping Stats Deploy Stage Modeling Data Production Data Variables Scoring Compare yโzโ
Agenda
Predictive Analytics:
Achieving Greater Decision Accuracy
ยป Capitalizing on predictive analytics
ยป Consider and establish value of predictive analytics to achieving strategic goals
ยป Develop focused strategy toward specifically targeted areas and needs