Session 63 PD, Annuity Policyholder Behavior. Moderator: Kendrick D. Lombardo, FSA, MAAA

Transcription

1 Session 63 PD, Annuity Policyholder Behavior Moderator: Kendrick D. Lombardo, FSA, MAAA Presenters: Eileen Sheila Burns, FSA, MAAA Kendrick D. Lombardo, FSA, MAAA Timothy S. Paris, FSA, MAAA

2 Timothy Paris, FSA, MAAA Ruark Consulting LLC Sessions 44 and 63 Annuity Policyholder Behavior May 17, 2016

3 Quick Tour of Annuity Industry Policyholder Behavior Experience Since

4 VA Surrenders by Company Surrender Rate Years Remaining in Surrender Charge Period

5 VA Surrenders by Year Surrender Rate Years Remaining in Surrender Charge Period

6 VA Surrenders by Guarantee Type Surrender Rate 7 or more or more Years Remaining in Surrender Charge Period GMIB GLWB GMWB GMAB None

7 VA Surrenders by Guarantee Value ITM 100+% ITM % ITM 25-50% Surrender Rate ITM 5-25% ATM OTM 5-25% OTM 25+% 7 or more or more Years Remaining in Surrender Charge Period

8 VA and FIA Surrenders Surrender Rate 7 or more or more Years Remaining in Surrender Charge Period VA GLWB VA No LB FIA GLWB FIA No LB

9 VA Surrenders GMIB Shock Lapse Surrender Rate 1Q 08 2Q 08 3Q 08 4Q 08 1Q 09 2Q 09 3Q 09 4Q 09 1Q 10 2Q 10 3Q 10 4Q 10 1Q 11 2Q 11 3Q 11 4Q 11 1Q 12 2Q 12 3Q 12 4Q 12 1Q 13 2Q 13 3Q 13 4Q 13 1Q 14 2Q 14 3Q 14 4Q 14 ATM <25% ITM 25%-50% ITM 50%-100% ITM

10 VA Surrenders GLWB Shock Lapse Surrender Rate 3Q 09 4Q 09 1Q 10 2Q 10 3Q 10 4Q 10 1Q 11 2Q 11 3Q 11 4Q 11 1Q 12 2Q 12 3Q 12 4Q 12 1Q 13 2Q 13 3Q 13 4Q 13 1Q 14 2Q 14 3Q 14 4Q 14 ATM <25% ITM 25%-50% ITM 50%-100% ITM

11 VA Surrenders and Partial Withdrawals Interaction Surrender Rate 7 or more or more Years Remaining in Surrender Charge Period No Prior WDs LT Full WDs Full WDs Excess WDs

12 VA Partial Withdrawal Frequency Withdrawal Frequency < Attained Age GLWB Non-Qual GLWB Qual No LB Non-Qual No LB Qual

13 VA Partial Withdrawal Frequency Withdrawal Frequency Duration

14 VA Partial Withdrawal Frequency Withdrawal Frequency < Attained Age Commencement Non-Qual Commencement Qual Continuation Non-Qual Continuation Qual

15 VA Partial Withdrawal Amounts Annual Withdrawal as % of Account Value < Attained Age GLWB No Living Benefits

16 VA Partial Withdrawal Amounts Excess Less Than Full Full

17 VA Mortality Actual / Expected Less rich Death benefits Living benefits Aggregate More rich

18 7 Lessons 17

19 Cleanse Craftsmen Range Speed Explore Change Credibility

20 Great Craftsmen Use Great Tools 19

21 Traditional Analysis Statistical Techniques

22 y x E(y x) Classical Linear Modeling g[e(y x)] Generalized Linear Modeling (GLM) Flexible framework Non-normal Non-constant variance Simple Linear Modeling

23 GLM Example Logistic Regression for Binary Response Variables ln pp 1 pp = ββ 0 + ββ ii xx ii Readily applicable to mortality and surrenders, where there are multiple simultaneous factors of influence 22

24 Model and Factor Selection Art + Science = Where you earn your pay Goodness of fit to historical data vs Predictive power for the future Many metrics and diagnostics 23

25 Aikake s Information Criterion (AIC) = 2k 2ln(L) Rewards goodness of fit (max likelihood function L), but with penalty for more model factors (k) to mitigate risk of overfitting the model on train data. Use to compare relative quality of alternative models. Lower is better. 24

26 Good ol Actual-to-Expected Ratios (A/E) Develop E using train data, compare to A from test data Examine in aggregate, by cohorts, and over time Look at range of outcomes and tails 25

27 Expert Judgment Business context, sensibility, materiality, parsimony More data usually beats more complex models Let the data speak Use simples models for complex data, and complex models for simple data 26

28 Logistic Regression Multiple Factors for VA Surrenders AIC Score All YrsRem +LB +LB/ITM +LB/ITM/Qtr +LB/ITM/Qtr/Size Factors

29 120% Logistic Regression Model Comparison for VA Surrenders 115% 110% 4-factor 5-factor Actual / Expected 105% 100% 95% 90% 85% 80% Expected Surrender Deciles

30 Ruark - Sample Logistic Regression Model for Policyholder Behavior (Surrenders) Factor Coefficient Std Dev e^coefficient Intercept YrsRemSurrChg U ITM DistChannel Example: consider a policy with YrsRem = 1, ITMGrp = 25, Dist = 2 B = = Odds = e^b = Probability of Surrender = Odds/(1+Odds) =

31 What s So Great About It? Goes beyond the endless series of reactionary point estimates to quantify range of behavioral values Consistent mathematical framework for assumption setting and updates Allows for company-level customization from max data set (industry) 30

32 Industry Data Traditional Analysis Statistical Techniques Expert Judgment

33 Optimization hinders evolution. Alan Perlis Statistics are no substitute for judgment. Henry Clay 32

34 Timothy Paris Chief Executive Officer Ruark Consulting LLC

35 2016 SOA LIFE & ANNUITY SYMPOSIUM Annuity Policyholder Behavior Eileen S. Burns, FSA, MAAA May 17, 2016

36 Notes from the field in predictive modeling What you want to know 1949 What you already know Data you use Data you could use What you can learn What you can act on significant significant 2 May 17, 2016

37 Case study 3 May 17, 2016

38 VA industry lapse assumptions: Base lapse Policy duration Base lapse Years 1 to 4 1.4% Years 5 to 7 3.4% Year 8 (shock) 16.8% Years 9 to % 18% 16% 14% 12% 10% 8% 6% 4% 2% 0% Base lapse Policy year 4 May 17, 2016

39 VA industry lapse assumptions: Dynamic Moneyness (BB/AV) Dynamic lapse factor % % % % % % % % 140% 120% 100% 80% 60% 40% 20% 0% Dynamic Lapse Factor In-the-moneyness 5 May 17, 2016

40 Caveats and Limitations A simple average of lapse assumption from multiple companies is very crude Companies in the Milliman survey and the predictive modeling data are not exactly the same Product features and distribution channels are ignored Don t take the numbers here and use as your own assumption 6 May 17, 2016

41 Industry Assumption vs. Experience Lapse Rates by Duration GLWB with 7-Year Surrender Charge Period Annualized Lapse Rate 14.0% 12.0% 10.0% 8.0% 6.0% 4.0% 2.0% 0.0% Years 1 to 4 Years 5 to 7 Year 8 Years Number of Records in Dataset (Millions) Policy Duration Aggregate Predicted Lapse Rate - Representative Industry Model Aggregate Actual Lapse Rate - Experience Data 7 May 17, 2016

42 Industry Assumption vs. Experience Observations: During SC, actual slightly higher than expected Actual shock lapse lower than expected Actual post-sc lapse lower than expected Potential sources for disconnect: Assumption is the product of base lapse and dynamic factor Post SC likely to be deeper ITM Dynamic lapse factor not low enough post SC? Are policyholders more efficient than expected in the industry lapse assumption? 8 May 17, 2016

43 Breaking lapse experience into components Lapse Rates by Duration GLWB with 7-Year Surrender Charge Period Annualized Lapse Rate 14% 12% 10% 8% 6% 4% 2% 0% Years 1 to 4 Years 5 to 7 Year 8 Years Number of Records in Dataset (Millions) Policy Duration Aggregate Predicted Lapse Rate - Predictive Model Aggregate Actual Lapse Rate - Experience Data 9 May 17, 2016

44 Component: Base Lapse Rates Duration Industry Base Lapse Rate Base Lapse Rate - Predictive Model Base Lapse Rate - Actual ATM experience Years 1 to 4 1.4% 1.3% 1.4% Years 5 to 7 3.4% 2.3% 2.4% Year % 9.6% 10.5% Years % 4.8% 5.1% Industry base lapse assumption appears too high at end of and after SC How is base lapse assumption set? Rely on base VA contract with no GLWB? 10 May 17, 2016

45 Component: Base Lapse Rates Riderless base VA experience v. VA with GLWB 20% 18% 16% 14% 12% 10% 8% 6% 4% 2% 0% Actual Data With Living Benefit ATM Actual Data Without Living Benefit VA with GLWB base (ATM) experience lower than riderless VA Introduction of GLWB changes policyholder behavior 11 May 17, 2016

46 Component: Dynamic lapse multiplier 200% GLWB Dynamic Lapse Behavior Dynamic Lapse Factors 100% 0% 75% 100% 125% 150% In-The-Moneyness (BB/AV) GLWB Baseline Predictive Model GLWB Industry Average - Representative Industry Model 12 May 17, 2016

47 Industry Assumption v. Experience Observation: Actual post-sc lapse lower than expected Conjectures: Assumption is the product of base lapse and dynamic factor Post SC likely to be deeper ITM Dynamic lapse factor not low enough post SC? Policyholders more efficient than expected in the industry lapse assumption? Findings: Industry experience suggests that base lapse assumption may be too high Suggests policyholders are not as efficient as expected 13 May 17, 2016

48 Conclusions Companies seem to be pretty conservative in setting dynamic lapse factors Actual experience still shows that lapse can be lower than expected Perhaps the conservatism in dynamic lapse is more than offset by the higher than expected base lapse Is looking at base and dynamic adjustment separately appropriate? In reality, policyholders make decisions based on a complicated network of lots of factors (including ITM) Lapse deserves a more rigorous and statistical analysis than it has been 14 May 17, 2016

49 Notes from the field in predictive analytics What you want to know 1949 What you already know Data you use Data you could use What you can learn What you can act on significant significant 15 May 17, 2016