Predictive Modeling in Life Insurance: Applications in In-Force Management. Aaron Sarfatti

Transcription

1 Predictive Modeling in Life Insurance: Applications in In-Force Management Aaron Sarfatti

2 PREDICTIVE MODELING IN LIFE INSURANCE APPLICATIONS IN IN-FORCE MANAGEMENT NOVEMBER 15, 2016 Aaron Sarfatti, Partner

3 CONFIDENTIALITY Our clients industries are extremely competitive, and the maintenance of confidentiality with respect to our clients plans and data is critical. Oliver Wyman rigorously applies internal confidentiality practices to protect the confidentiality of all client information. Similarly, our industry is very competitive. We view our approaches and insights as proprietary and therefore look to our clients to protect our interests in our proposals, presentations, methodologies and analytical techniques. Under no circumstances should this material be shared with any third party without the prior written consent of Oliver Wyman.

4 Industry perspectives predictive analytics as an in-force management tool Why are insurers looking at in-force management? Extract value from in-force blocks Increased attention among life insurers driven by: Capital markets headwinds Limited organic growth potential Sometimes, a desire to recycle capital from legacy businesses How do predictive analytics support inforce management? Address critical in-force management questions: What actions will the policyholder or agent take that affect policy value? What methods are most effective to enhance/alter these actions? How do in-force predictive models differ from actuarial assumption models? Typically a short-term event being predicted (but with long-term implications) More use of event data - life changes, recent policy actions, admin data 2

5 Questions to address in a typical in-force management model Life in-force retention example What drives the loss? Reason for lapse, % How effective are saves? Save rate (% retained) What are combined results? Lapse experience by reason Date: Q Save rate by protocol Date: Q Outcomes Date: Q Lapse reason 1M avg. 3M avg. No action 15% 1 Affordability canceling cover % 35% Cheaper price 30% 20% Better product 10% 10% No insurance need / life-stage 20% 30% Job change 5% 5% Other 5% 0% Send letter A to agent Send letter B to policyholder Send letter C to policyholder Call policyholder with script A Other 27% 28% 39% 32% 48% Restored premium 39% Issued rider 6% Policy lost 13% No policy change 32% Reduced face value 10% Drives both retention efforts and customer management Enables optimization of existing retention efforts Provides more complete picture on retention cost and effectiveness 3

6 What types of predictive models are used? Life in-force retention example Data required Event predictions Examples Required analytics Augmented experience study data Premium patterns Partial withdrawals Address changes Classify premium payments into patterns and identify deviations Determine lapse / surrender rates by pattern and deviation types Determine lapse / surrender rates and timeframe by partial withdrawal amounts Determine loss rates by reported address change category and timing Producer data Orphan policies Determine loss rates by orphan type and timeframe External data Life events Identify how events (marriage/divorce, job loss, credit changes, etc.) affect policy loss and timeframe Retention models often require networks of inter-connected predictive models 4

7 How is a typical retention predictive model structured? Common model structure Differences from usual actuarial assumption predictive models Baseline risk Event triggers Greater use of short term trigger data Inputs Duration Attained Age Premium Pattern Premium stoppage Premium reduction Often complements long term assumption data Model Baseline risk Trigger A risk Trigger B risk Similar usually a GLM model form Complements traditional predictive model to inform actuarial assumptions Filters Policy attribute e.g. face amount Policy attribute e.g. funded status Use of attributes to exclude data Model designed on small sets of data Output Other policies Policies for retention Focus on subset of policies most likely to exhibit predicted action 5

8 Does model development & refinement vary from pure actuarial models? Test-and-learn approach central to efficient resource management Illustration of test-and-learn framework Quick-win identification Ingoing hypotheses based on industry experience will expedite opportunity identification Pilot Testing Ongoing monitoring and refinement to allow test and learn approach to pilot development High value & accessible Prioritized System Univariate Analysis Reason identification Pilot Design Pilot Implementation Performance Reports Univariate analysis Identification of lapse / surrender reasons to support design Robust tracking and reporting enables pilot design refinement Business engagement much greater, timeframes longer 6

9 Steps to execute an in-force management predictive model pilot application Development steps of lapse and surrender propensity models Construct data hub Build propensity models Multiple data sources consolidate into single location Integrate new data sources over time Define propensity models based on usage: Identification of at-risk policies Policy prioritization Assess modelling techniques Policy-level surrender propensity Date: Q Product: UL 5 Timeframe: months Type: Surrender Policy #00001 Policy #00002 Policy # % 27% 28% Define call scripts and / or print materials to enable pilots Policy # % Define pilots to apply models Review proposed pilots with broader business Train associates and specialists to support pilots Policy # % Integrate with prioritization Incorporate lapse, surrender, and replacement propensity into policy prioritization for ongoing retention efforts and pilots Train associates and specialists on new prioritization Policy # % 7

10 Data hub construction Important to start simple, build business engagement before large IT outlays Sequence of data hub construction Single data source Multiple unlinked sources Multiple linked sources Integrated environment Demonstrate value of predictive modeling with single trigger propensity models Begin conceptual development of multi-variate models Identify data field keys to piece together systems Expand trigger identification for simple propensity models Prioritize efforts to link data sources for multi-variate models Develop linked multi-system hub Verify lift in policy identification provided by more complex modeling Prove feasibility of fully integrated environment Continue to refine models using tracking and reporting from pilot implementation Retention protocols will require careful design due to lack of clear reason for policyholder contact Marginal gains from new data sources and complex modeling techniques quickly diminish Business judgement needed to determine appropriate stopping points for model and data hub 8

11 How should model performance be evaluated? Requires traditional statistical and overall business impact evaluations Traditional statistical evaluation (ensures models adequate) Business impact evaluation (for overall effect) 100% Premium restoration rates True positive rate 90% 80% 70% 60% 50% 40% 30% Rate (% policies) Cumulative surrender rates Retention uplift Pilot group Control group Time since premium deviation 20% Long-term effect 10% 0% 0% 20% 40% 60% 80% 100% False positive rate Model 1 (ROC = 0.823) Model 2 (ROC = 0.846) Model 3 (ROC = 0.808) Model 4 (ROC = 0.852) Rate (% policies) Intervention spike Pilot group Control group Time since premium deviation Base (ROC = 0.50) 9

12 Conclusions Conclusion Predictive models often similar in structure, data as traditional assumption models Comments GLM approach common Often works in tandem with assumption models Model development a relatively small part of the challenge Significant time devoted to: Data identification and organization Pilot design, hypothesis generation, etc Business engagement and demonstrating success more critical Businesses will be skeptical it s new, complex and budgets are tight Patience is a virtue Need to find positive signals P&L benefit not immediate 1-2 year evaluation timeframes necessary 10

13 QUALIFICATIONS, ASSUMPTIONS AND LIMITING CONDITIONS This report is for the exclusive use of the Oliver Wyman client named herein. This report is not intended for general circulation or publication, nor is it to be reproduced, quoted or distributed for any purpose without the prior written permission of Oliver Wyman. There are no third party beneficiaries with respect to this report, and Oliver Wyman does not accept any liability to any third party. Information furnished by others, upon which all or portions of this report are based, is believed to be reliable but has not been independently verified, unless otherwise expressly indicated. Public information and industry and statistical data are from sources we deem to be reliable; however, we make no representation as to the accuracy or completeness of such information. The findings contained in this report may contain predictions based on current data and historical trends. Any such predictions are subject to inherent risks and uncertainties. Oliver Wyman accepts no responsibility for actual results or future events. The opinions expressed in this report are valid only for the purpose stated herein and as of the date of this report. No obligation is assumed to revise this report to reflect changes, events or conditions, which occur subsequent to the date hereof. All decisions in connection with the implementation or use of advice or recommendations contained in this report are the sole responsibility of the client. This report does not represent investment advice nor does it provide an opinion regarding the fairness of any transaction to any and all parties.