How Advanced Pricing Analysis Can Support Underwriting by Claudine Modlin, FCAS, MAAA

What Is Predictive Modeling Predictive modeling uses historic data to identify patterns that can be used to predict future behavior. At individual risk level And also aggregated to class or portfolio level These predictions are then used to inform business decisions. Insurance predictive modeling takes many forms. Costs Conversion/renewal behavior Operations Wider customer behavior 4

The Problem With Prediction It is exceedingly difficult to make predictions, especially about the future. (Niels Bohr, Nobel laureate in physics) It is often said there are two types of forecasts... lucky or wrong. (Institute of Operations Management ) I think there is a world market for maybe five computers. (Thomas Watson, Chairman of IBM, 1943) 640K ought to be enough for anybody. (Bill Gates in 1981 maybe) 5

The Problem With Prediction Algorithms priced book at $23.6m on Amazon 6

But In 1999, Billy Beane (manager of the Oakland Athletics) found a novel use of predictive modeling. Using predictive analytics, Beane was able to identify and hire players undervalued by the market. A year later, the A s ranked 2 nd. Implication You only need to be more accurate than before. Predictive models can aid the human expert BUT need the expert to define the framework. 7

Applications Within Insurance Claims and claims analytics Finance Enterprise risk management $ Reserving Customer services Analytics Team Underwriting & pricing Distribution Marketing Planning Predictive modeling is being used to help integrate all aspects of companies operations and identify the true customer value 8

Predictive Modeling Use Has Increased for Nearly Every Line of Business; Standard Commercial Carriers Have Aggressive Plans, but a Gap Will Continue in Specialty Lines Does your company group currently use, or plan to use, predictive modeling in underwriting/risk selection and/or rating/pricing for the following lines of business? (Q.2) Personal automobile Homeowners Workers compensation General liability/cmp/bop Commercial property/cmp/bop Commercial automobile Specialty lines 2013 (n = 40) 2012 (n = 45) 2013 (n = 37) 2012 (n = 41) 2013 (n = 34) 2012 (n = 27) 2013 (n = 45) 2012 (n = 33) 2013 (n = 47) 2012 (n = 34) 2013 (n = 42) 2012 (n = 33) 2013 (n = 31) 2012 (n = 14) 26% 32% 30% 28% 24% 13% 30% 80% 75% 62% 58% 47% 38% 33% 40% 15% 5% 19% 6% 33% 5% 33% 9% 41% 12% 41% 21% 45% 22% 34% 53% 15% 38% 32% 60% 12% 43% 33% 45% 22% 42% 48% Currently use Plan to use Do not use and no plans to use Personal lines Standard commercial lines Specialty commercial lines Source: Towers Watson s 2013 Predictive Modeling Survey Base: U.S. respondents giving a valid answer (percentages exclude Do not write this line of business ). 9

While All Carriers See Highly Favorable Bottom-Line Results, Commercial Carriers See More Impact on Top-Line Growth What impact has predictive modeling had in the following areas? (Q.19) Rate accuracy PL CL 85% 96% 5% 10% 4% Profitability PL CL 80% 78% 15% 18% 5% 4% Bottom line Loss ratio improvement PL CL 80% 74% 15% 22% 5% 4% Expansion of underwriting appetite PL CL 45% 48% 50% 48% 5% 4% Renewal retention PL CL 25% 52% 55% 39% 20% 9% Top line Market share PL CL 35% 39% 45% 52% 20% 9% Positive impact No impact, neither positive nor negative Negative impact Source: Towers Watson s 2013 Predictive Modeling Survey Base: U.S. respondents currently using predictive modeling for at least one line of business (personal lines n = 20, commercial lines n = 23). 10

Risk Selection A variety of considerations come into play when making a risk selection decision: Underwriting appetite and goals Underwriting guidelines Policy terms and conditions Expected cost of the risk Expected profitability of the risk Confidence around that profitability estimate The longer-term profit potential of the whole account (i.e., customer lifetime value) This presentation will address how analytics support these aspects. 13

Expected Cost of the Risk Risk models predict the individual risk s expected loss cost. Separately by claim type (e.g., property vs. liability) Typically model frequency and severity (and combine to loss cost); may also include a large loss propensity model Predictors in the model can include existing rating variables as well as new underwriting information (internal or external data) These models have been used in P&C insurance for over a decade. Started in personal lines (homogeneous risks, large volumes of clean data) Expanded to commercial and specialty lines Many companies are on third or fourth generation efforts 14

Risk Models Are Built in a Statistical Framework but Infuse the Business Acumen of the Modeler as Well Advanced Interaction Detection Model Validation 15

Need a Variety of Techniques to Solve Many Types of Business Problems We leverage a broad array of techniques. Examples include: Decision trees Systematic approach to slicing data Factor selection tool Spatial Analysis Define territories based on experience rather than geopolitical boundaries Define class codes using own experience These methods must be employed carefully to improve performance without losing the transparency of the model. 16 15

Selecting Risks Based on Cost Cost is one indicator of good vs. bad risk. High cost risks can be good risks, however, assuming the right premium charged, coverage conditions applied, etc. A better metric for risk selection is expected loss ratio. 17

Underwriting Profit Scoring Models By incorporating premium into risk models, they can be easily translated into underwriting profit scoring models that rank risks according to profitability underlying current rates. Raw Scorecard Loss Control Company Size Yes 0 Small 0 No 7 Medium 9 Large 6 Territory A 0 Policy Tenure B 6 New 12 C 9 1 7 D 13 2 4 E 16 3+ 0 Score < 20 20 25 26 30 31+ Cumulate scores to generate policy level scores Score Factor < 20 0.90 20 25 1.00 26 30 1.05 31+ 1.20 Derive score factors through modeling techniques 18

Typical Underwriting Profit Score Relativity = 3.0 Relativity = 0.35 Relativity = 1.0 Breakeven without score 19

Accept/Reject Scoring Accept Reject Relativity = 1.5 at Score = 40 20

Improving Manual Rates via Underwriting Tiers Tier #3 = 2.0 Tier #2 = 1.0 Tier #1 = 0.4 21

Providing Schedule Rating Guidance Credit 1.0 Debit Max 25% Credit Max 25% Debit 22

Tiering With Schedule Rating Tier #3 = 1.50 Tiering Schedule Rating 25% Credit Tier #1 = 0.50 25% Credit Tier #2 = 1.0 25% Debit 23

Comparison of Loss Ratio Improvement by Adoption of Predictive Modeling (Commercial Lines) 4.00% Change in Earned Loss Ratio PM For Pricing/Risk Tiering = YES PM For Pricing/Risk Tiering = NO 3.00% PM For Risk Selection = YES PM For Risk Selection = NO 2.00% 1.00% 0.00% -1.00% -2.00% -3.00% -4.00% 2012-2013 Source: Towers Watson s Commercial Lines Industry Price Survey (CLIPS) 2 nd quarter 2014 24

Considering Both Expected Loss Ratio and Associated Variability Can Improve Underwriting Process Flow Companies can establish underwriting processes in consideration of both the underwriting profit score and the degree of confidence in the score calculated for a particular risk. This allows underwriters to focus on those risks that need it most. Variability High Low Underwriting (drill down to increase confidence) (A3) Automated Underwriting (A1) Underwriting (apply loss prevention or other UW support) (A4) Automated Underwriting w/ Price Check (A2) Low High Expected Loss Ratio 25

Underwriting Process Flow In a recent investor day presentation, Travelers showed that between 2006 to 2013, the number of submissions for its Select Express product that were straight-through processed increased from 17% to 80%. 26

Selecting Risks Based on Expected Profitability Selecting risks based on expected loss ratio will result in a portfolio of risks that are immediately profitable. What about risks that may become profitable in the future? Longer-term profitability is measured as the net present value of the cash flows attributed to the relationship with a customer. This requires estimation of: A profit metric (revenue minus costs) How the risk will change over time (both customer-initiated changes and companyinitiated changes) Likelihood of persisting Likelihood of cross-selling and up-selling A discount rate Referral Cross-sell Up-sell Up-sell Referral Current Product 27

Customer Value Chain Policyholders make decisions throughout the lifecycle. Policyholder behavior can be modeled. Prospect Identification Lead Generation Lead Acquisition Customer Retention Up-Sell / Cross-Sell Lead Generation Conversion Retention Propensity 28

Drivers of Retention/Conversion Policyholder Characteristics Policyholder Attitudes Payment Payments Customer characteristics Coverage Vehicle Age Vehicle Group Efficiency o g ade roduct Market Competitiveness Minimum Limit External influences Regulation Competitor Acquisition Initiatives Mandatory Insurance ge ent od se or ate Product Holdings & Activity Payments Customer triggered changes Mid Term Adjustments Customer Relationship C C 29

Behavioral Analysis 30

Behavioral Analysis 31

Behavioral Analysis 32

Demand Curves Vary by Risk Demand Price 33

Elasticity Varies by Price Low Elasticity Retention Rate 0.5 1.0 High Elasticity 2.0 1.0 0.5 Price 34

Modeling Elasticity vs. Demand 35

Customer Lifetime Value (CLV) CLV is the cash flows attributed to the relationship with a customer. Whole of the Wallet Lifetime Value Expected Cross Sales Lifetime Value Acquisition Value Profit Potential Profit Time Future Lifetime Value Value at Risk 36

Calculating CLV: Simple Example Annual premium: $1,000 (constant over time) Profit: 10% accounted for at the end of the policy and fixed over the life of the customer Retention: 90% with no mid-year termination Acquisition cost: $400 Discounting rate: 15% CLV = 100/1.15 + 100 * (0.9) / 1.15 2 + 100 * (0.9) 2 / 1.15 3 +. 400 37

Uses of CLV Risk selection decisions Marketing Customer relationship management Provide preferential services Identify targets for retention campaigns Offer certain products to specific customers Broker management and compensation Future value High Low Nurture and cross-sell Maintain service level and reduce costs Low Current value Develop and acquire similar prospects Retain High 38

Leveraging Analytics for Distribution Management These metrics (or combinations) can be very useful in informing the distribution channel and influencing its performance. Expected cost Expected loss ratio Expected demand Customer lifetime value The following is a case study demonstrating this. Up-sell 40

Predictive Modeling Case Study in Agency Management Two Models Were Built 1.Demand: New Business Conversion Model Identifies the types of customer the company attracts 2.Risk: Expected Loss Cost Model (which can be easily converted to Expected Loss Ratio model): Identifies which customers are expected to be profitable Business was reviewed in total and separately by agent to determine which agents are generating the best business. 41

Demand and Loss Ratio by Agent 42

Expected Demand and Expected Loss Ratio by Largest Volume Agents 43

Expected Loss Ratio (71%); Expected Demand (63%) 44

Plot Expected Loss Ratio Case Study Profitable Unprofitable Low Expected Loss Ratio High 45

Plot Expected Demand Case Study Expected Demand Low High High demand Low demand Low Expected Loss Ratio High 46

Expected Loss Ratio and Demand Simultaneously Case Study Expected Demand Low High Best Customer Good Customer Adverse Selection Undesirable Low Expected Loss Ratio High 47

Expected Loss Ratio and Demand Across Company 48

Expected Loss Ratio and Demand Across Company 63% 71% 49

Expected Loss Ratio and Demand Across Company Best Customer Adverse Selection 63% Good Customer Undesirable 71% 50

Predictive Modeling Case Study Identifying Best Agents Simultaneously plot expected loss ratio and demand across each agent separately 51

Agent 19 Low Expected Loss Ratio and High Demand 52

Agent 19 Low Expected Loss Ratio and High Demand Best Customer Adverse Selection Good Customer Undesirable 53

Agent 2 Low Expected Loss Ratio and Low Demand 54

Agent 2 Low Expected Loss Ratio and Low Demand Best Customer Adverse Selection Good Customer Undesirable 55

Agent 13 High Expected Loss Ratio and High Demand 56

Agent 13 High Expected Loss Ratio and High Demand 57

Agent 18 Low Expected Loss Ratio and Low Demand 58

Agent 18 Low Expected Loss Ratio and Low Demand Concentration of risks is desirable. Adverse selection on a small number of policies is offsetting the profitability on the majority of business. 59

Integrating Demand in Pricing Decisions Elasticity Analysis: If we can understand both expected cost and the customer s sensitivity to price offers, we can develop an optimal strategy that achieves the company s objectives in terms of profitability, volume and other enterprise goals. Market Competition Current Portfolio Needed Price Change Future Portfolio Future Portfolio 61

Implementing Price Decreases Example 1 Account A Account B Actuarial Cost Model Indication -15% change -15% change Elasticity High Low CRM Strategy Price-based Service-based Near-Term Pricing Action Change rate -15% Change rate -5% Results Attract/retain more business Maintain strong profit margin 62

Implementing Price Increases Example 2 Account C Account D Actuarial Cost Model Indication +15% change +15% change Elasticity High Low CRM Strategy Price-based Service-based Near-Term Pricing Action Change rate +5% Change rate +15% Results Attract/retain more business Rapidly restore profitability 63

Selecting Optimal Prices In regulated lines that require a filed rate to be charged (e.g., private passenger auto) the filed rate is often selected as a cost-based price with prudent deviations for: Business objectives: competitive position, profit and volume goals, tolerance for dislocation Constraints (regulatory, operational) Deviations can be determined based on: What if scenario testing Optimization algorithms This concept of using analytics to select prices that better achieve goals is completely relevant to commercial lines as well ( more on that later). 64

Quantitative Competitive Analysis Can Advise Where Price Changes Can Improve Competitive Position 65

Selecting Optimal Prices Many other key performance indicators (KPIs) can be measured in order to make sound pricing decisions. Profit $ or % Volume (premium volume, policies in force, retention rate, conversion rate) Subsidy Dislocation Competitive position This requires data integration in a price assessment environment A customer dataset (e.g., in-force policies) Current premiums Competitor premiums Risk models Demand models Expense assumptions The following case study considers what if scenario tests of changing a filed rate order calculation. 66

Price Assessment Illustration 67

Price Assessment Illustration 68

Price Assessment Illustration 69

Price Optimization Modern analytical approaches (e.g., price optimization) use mathematical algorithms to systematically integrate risk and policyholder behavior in selecting a price for the individual risk. This removes judgmental bias from the selection process. 70

Customer Knowledge Is Used to Simulate a (Constrained) Search Space of Potential Price Changes A search space for one policy constrained by allowable rate change S1 S2 S3 S4 Expected profit Different business objectives have different optimal prices Retention rate Profit per policy S1 Maximize volume S2 Current rates S3 Cost-based rates S4 Maximize profit 400 410 420 430 440 450 460 470 480 490 500 510 520 530 540 550 560 570 580 590 600 Price Premium for policy in question 71

Search Spaces Are Simulated for Each Individual Policy 85% 80% 127 92% 70 Retention 75% 70% 65% 60% 55% 122 117 112 Expected Discounted Contribution Retention 90% 88% 86% 84% 82% 60 50 40 30 Expected Discounted Contribution 50% 107 80% 20 45% 102 290 310 330 350 370 390 410 430 Proposed Premium 78% 10 180 190 200 210 220 230 240 250 260 270 Proposed Premium Retention Expected Discounted Contribution Retention Expected Discounted Contribution 90% 83% 125 88% 85 120 Retention 86% 84% 82% 80% 78% 76% 75 65 55 Expected Discounted Contribution Retention 78% 73% 68% 115 110 105 100 95 90 Expected Discounted Contribution 74% 45 63% 85 72% 80 70% 35 200 210 220 230 240 250 260 270 280 290 300 Proposed Premium 58% 280 300 320 340 360 380 400 Proposed Premium 75 Retention Expected Discounted Contribution Retention Expected Discounted Contribution 72

Explore the Simulated Search Space to Identify Potential Price Scenarios ILLUSTRATIVE Optimization identifies and summarizes a range of options in the search space it is not just about increasing profit. 3,500 Total Expected Discounted Contribution Period 1 ( '000) 3,000 2,500 Profit 2,000 1,500 1,000 Cost Based Current 500 48,000 49,000 50,000 51,000 52,000 53,000 54,000 55,000 56,000 Volume Total Retention Period 1 1 year simple contrained 73

A Wide Array of Constraints Are Introduced Into the Optimization Algorithm, Reflecting Internal and External Considerations Policy-level constraints (e.g., an allowable range of premium changes) Portfolio level constraints Dislocation Profitability Competitiveness Universal constraints The result of the optimization is expressed as a rate order calculation. The rate should not be inadequate, excessive or unfairly discriminatory. 74

Optimization Can Be Expanded to Consider Profit, Volume and Other Metrics (e.g., Customer Lifetime Value, Account Profit) 75

Uses of Price Optimization Price optimization is gaining speed in personal lines with demonstrated improvement in profit and/or volume. In unregulated commercial lines, producing a suggested retail price (via price optimization) can be used for: Risk selection Informing agent/underwriter judgment Guiding final price decisions 76

Summary Insurance companies are competing on data and analytics. Arming the underwriter and distribution channel with more insights leads to better decisions. Analytics can help estimate loss costs, profitability, customer behavior and customer lifetime value. Integrating these analyses can support underwriters with: Risk selection and underwriting process flow Distribution management Guiding price selections 77

Contact Information Claudine Modlin, FCAS, MAAA Director Towers Watson San Diego, CA (805) 499-2164 Claudine.Modlin@ 78