Session 113 PD, Data and Model Actuaries Should be an Expert of Both. Moderator: David L. Snell, ASA, MAAA

Transcription

1 Session 113 PD, Data and Model Actuaries Should be an Expert of Both Moderator: David L. Snell, ASA, MAAA Presenters: Matthias Kullowatz Kenneth Warren Pagington, FSA, CERA, MAAA Qichun (Richard) Xu, FSA SOA Antitrust Disclaimer SOA Presentation Disclaimer

2 Richard Xu, PhD, FSA VP & Actuary, Head of Data Science RGA Oct 2016

3 Add Model to Your Data and Business Richard Xu, PhD, FSA VP & Actuary, Head of Data Science RGA Oct

4 The Problem Current UW for life is expensive & time-consuming needs lab tests (blood/urine), paramedics (BMI, BP) takes weeks, decreasing take-up rate Internal Demands Compete in existing and new markets Increase efficiency Develop new products Be an expert in new data and uses Protect mortality slippage External Demands Enrich customer experience Decrease cycle time Don t ask too many questions, use data Place more cases Improve the underwriting process

5 The Solution Expertise Data Science Technology Underwriting Mortality Reinsurance Data Driven UW Capabilities Data & Analytics Underwriting Rules Automated Underwriting Predictive Models

6 Current Industry Trend Use model to select risks Combine rule-based approach with model results Rely on various data sources to make UW decision Use traditional UW for high risks

7 New Problem Current industry standard being preferred + smoking status No medical & lab test; how about smoking? Smoking status - the single most important risk factor after age/gender; about 200% mortality difference Normally detected by medical test ~20% lies under traditional UW; vary by channel What possibly could happen under new UW Bad news: more lies on SM/NS; more anti-selection Good news: still honest applicants 6

8 Any Solution? Can we use data-driven solution to differentiate SM vs. NS, without medical test? We can not directly observe SM/NS, so what are we inexplicitly assuming? Smoking status may be related to indirect medical info, behavior data, past application history? Causation vs. correlation Possible data sources? Application, especially SM/NS question Prescription data Motor vehicle records MIB data Other 3 rd party data, what s available 7

9 Model Specification Do we expect a simple pattern in the data? Simple model, e.g. GLM, may not be effective We may need more sophisticated algorithms Neural network, random forest, SVM, boost, etc. Even better, ensemble approach Even simple average of models is better There are more sophisticated approach Can actuaries do it? Absolutely, with education background and understanding of data and business but need certain training in modeling 8

10 Model Performance A very sophisticated model to predict smoker vs. nonsmokers 100% 80% Smoking Model Lift Curve 60% 40% SM Average 20% 0%

11 Implementation Self-claimed nonsmoker SM/NS Model Prediction Predicted nonsmoker Fast UW Process Form Application Form Predicted nonsmoker Self-claimed Smoker Full UW Process 10

12 Application SM/non-SM Model can be used for triage purpose Non-smokers can go to data-driven UW Smokers are required to go to full UW Why triage, not UW? FCRA issue Very high pass thought rate for non-smoker No lab test required for the process Actuaries can further analysis cost/benefit & find optimal parameters Analysis of cost saving from no lab test, and additional mortality cost due to prediction uncertainty This calculation could be a function of age/gender/face amount Fine a optimal point where lab-test savings > mortality slippage 11

13 Considerations for actuaries Data / business Data is always #1 issue in data-driven application Domain knowledge is a key in modeling process Modeling Weak spot in actuary skill sets May need more education & training SOA is working on solutions; it is coming Actuaries can not miss the opportunity Solid statistics foundation & basic modeling skills Already expert of data & business; unique position Statistic analysis is the future, and you could be expert on both data & modeling 12

14 Add Model to Your Data and Business Richard Xu, PhD, FSA VP & Actuary, Head of Data Science RGA Oct

15 Matthias Kullowatz Modeling Techniques October 25 th, 2016

16 Non-parametric models 2

17 Examples Decision Tree-based Gradient-Boosted Machine Random Forest Generalized Additive Model Neural Network Support Vector Machine 3

18 Gradient-Boosted Machine Composed of many decision trees Each tree is fit iteratively to improve upon previous trees Flexible variable types allowed 4

19 Utilizing a GBM Guide variable selection Identify non-linear effects Discover variable interactions 5

20 Binary Response 6

21 Parametric model options Poisson GLM λ = ee β 0+β 1 xx 1 + +β kk xx kk Logistic GLM (binomial) π = eeβ 0+β 1 xx 1 + +β kk xx kk 1 + ee β 0+β 1 xx 1 + +β kk xx kk LLLLLLLL = YY ii ln( π) + (1 YY ii ) ln(1 π) 7

22 Questions How likely is it that a given policyholder will lapse in the coming period? will start lifetime withdrawals in the coming period? die in the coming period? What is the probability that the Cavs will repeat as NBA champions? 8

23 GBM Example Variable importance: Covariate Relative Importance Moneyness 31.6 Duration 26.5 SC Phase 18.9 Attained Age 14.6 Garbage 4.6 Distribution Channel 3.8 9

24 GBM Example Non-linear effects of moneyness: 10

25 Survival Response 11

26 Parametric model options Cox Proportional Hazards HHHHHHHHHHHH MMMMMMMMMMMMMMMMMMMM = ee β 0+β 1 xx 1 + +β kk xx kk Accelerated Failure Time (regression) θθ = ee β 0+β 1 xx 1 + +β kk xx kk LLLLLLLL = dd ii ln ff(xx ii θθ) + (1 dd ii ) ln SS xx ii θθ 12

27 Questions How long before a given policy holder dies? makes a lifetime withdrawal? lapses her policy? How many years do I have to wait to see the Mariners in the World Series? 13

28 Mortality Non-linear effects of BMI: 14

29 GLWB Withdrawal Interaction between issue age and qualified status: 15

30 Takeaways Non-parametric models have the capability to inform the construction of parametric models. The natural format of the GBM lends itself to intuitive interpretations of: variable importance non-linear relationships interactions 16

31 Thank you 17

32 Ken Pagington 113 PD Data and Model - Actuaries Should be an Expert of Both October 25, 2016

33 Goal Promote behavior change Introduce a filter for where to focus your time Define how one might iterate through data/model expertise Provide examples of little things causing big problems Discuss the benefits of proactively partnering with insurance operations (finance and non-finance) 2

34 Your mission if you choose to accept: Mission Statement - Past, Present, and Future Past: Understand and use data from the past Future: Create assumptions and models projecting anticipated futures Present: Use both to make decisions in the present! Projections Questions to ask yourselves about data/models you use: Model Building Policyholder Behavior How will this work be used to make decisions? Experience Analysis Decision- Making Assumption Building Could I do something more refined if I poked around some more? What human beings touch this process? 3

35 Be a student of the business and find a balance of data and model Data GIGO Captured, but not used Cleansing vs Altering Model Model what we can t measure? Fudge Factor Aggregation vs Granularity One Approach to finding the right balance is to follow the money, iteratively: Make as detailed assumptions & model as you can or is appropriate Forecast with a model Compare actuals vs forecast and explain results regularly Material unexplained variances should lead back to the hunt for data, partner with operations, build new assumptions, and then back to the model Back-cast history to validate new info Keep forecasting ahead and explaining results (repeat cycle as needed) 4

36 Sometimes, the value is in the details. Experts in data & model are capable of finding and solving problems before they surface. Examples 5

37 Example 1: Inforce + Sales = Premium Forecast Process Extract inforce data from database Get sales assumptions from planning area / field support Modeling area creates a Premium forecast Pause to consider What could possibly go wrong? What questions would you ask? Whom would you seek out for information? 6

38 Example 1: Non-Trivial Pitfalls Process Pitfall Impact to Forecast Extract Inforce Sales Forecast Sales & Inforce interaction Forecast Bias Model Policy issue timing Model not seriatim Admin system processes policies early and the inforce extract contains future effective dates that haven t paid Sales area reports submitted sales, but not every client pays, so there is always less actual Premium The inforce extract overlaps with some sales Is there a reason for Sales area to over/understate sales forecast? Assumes 1 st of month issue Does not reflect policy duration impact of lapse Overstated at the start Overstated throughout forecast horizon Overstated at the start Could be biased high or low Overstated throughout prediction Could be biased high or low 7

39 Example 2: Profitable Reinsurance Process New business model runs each quarter s worth of business This quarter shows a lift in profitability, but not sure why at first Analysis shows lift from reinsurance, but need to dig more for more information Pause to consider What could possibly go wrong? What questions would you ask? Whom would you seek out for information? 8

40 Example 2: Different assumptions to blame Reinsurance Collected Reinsurance file read in by Model is different. It contains higher substandard mortality. Inforce extract does not contain the same substandard mortality ratings. Death benefit projection is much lower. Is there a reason the substandard ratings differ? Which file is right? What does the reinsurance system show? What does the admin system show from Underwriting? Death Benefits Paid Out 9

41 Summary

42 Summary Think about data=past, model=future, and decisionmaking=present Define what decisions will be made based on data/models before investing time or effort Set up a regular process to explain results Begin conversations with insurance operations to promote line-of-sight into each others work Uncover new ways you can support the business 11

43