• No results found

Session 35 PD, Predictive Modeling for Actuaries: Integrating Predictive Analytics in Assumption Setting Moderator: David Wang, FSA, FIA, MAAA


Academic year: 2021

Share "Session 35 PD, Predictive Modeling for Actuaries: Integrating Predictive Analytics in Assumption Setting Moderator: David Wang, FSA, FIA, MAAA"


Loading.... (view fulltext now)

Full text


Session 35 PD, Predictive Modeling for Actuaries: Integrating Predictive Analytics in Assumption Setting


David Wang, FSA, FIA, MAAA


Guillaume Briere-Giroux, FSA, MAAA Eileen Sheila Burns, FSA, MAAA


Presented by

Eileen S. Burns, FSA, MAAA


October 27, 2014

Integrating Predictive Modeling in

Assumption Setting


What makes predictive modeling different?

 Results

– More granularity – Quantifiable goodness-of-fit

 Foundations

– Different tools – More data prep





Human and physical capital

 Decision #1: Project staff

– Actuary – Statistician

 Supporting staff

 Decision #2: Program

– Comfort – Flexibility – Processing efficiency – Cost

 Computing resources

OR ?





OR ? http://commons.wikimedia.org/wiki/File:SAS_logo_horiz.svg http://commons.wikimedia.org/wiki/File:R_logo.svg


Exploratory data analysis

 Programmer, statistician, and actuary

 Data visualization tools

http://commons.wikimedia.org/wiki/File:SAS_logo_horiz.svg http://commons.wikimedia.org/wiki/File:R_logo.svg http://commons.wikimedia.org/wiki/File:Tableau_Software_Logo_Small.png


How is each field distributed? Is it consistent over time?

Which variables are related to the target variable, and how?

Data visualization – on raw and prepped data

Policy value Policy value Category Category Time Time


Example 1: Gender

Problem: Missing

extreme number of

genders, male and

female counts


Solution: Code

problem referencing

secondary insured

by mistake (gender2

vs. gender field)



Example 2: Distribution channel

Problem: Large

number of distribution

channels missing

Solution: No

change. Discovered

issue by looking

across time. Fix is

only available for

policies that don’t

lapse, and would

distort experience.


Example 2: Distribution channel

Problem: Large

number of distribution

channels missing

Solution: No

change. Discovered

issue by looking

across time. Fix is

only available for

policies that don’t

lapse, and would

distort experience.

08 Q2 08 Q3 09 Q1 08 Q4 09 Q2 09 Q3 09 Q4 10 Q1 10 Q2 A B C NA


Example 3: Policy status records

08 Q2 08 Q3 09 Q1 08 Q4 09 Q2 09 Q3 09 Q4 10 Q1 10 Q2

Problem: No deaths

in any Q4

Solution: No

change. A data issue

with no feasible

solution. Only affects

scaling and only for a

small percentage of



Example 4: Rider prevalence

08 Q2 08 Q3 09 Q1 08 Q4 09 Q2 09 Q3 09 Q4 10 Q1 10 Q2

Problem: no other

rider features seen in

one quarter.

Solution: Analyzed

prior and subsequent

quarters to identify

policies with riders.

Necessary to avoid

modeling policies

with riders as no-rider



Example 5: Assessing policy values

Death benefit

Log distribution

Log death benefit

Relative distribution DB/AV % of dist <= 1.00 25% 1.01 - 1.25 50% 1.26 - 2.00 10% 2.01 – 5.00 5% 5.01 – 9.00 5% 9.01+ 5%

Problem: What is a



Solution: Check log

distribution, and

check relative to

similarly-scaled field.

Led to recognizing

some $4$ benefits,

and some DB coded

as increase over AV.


Example Recap

 Response variable

– Policy status

 Factor variables

– Rider prevalence – Gender – Distribution channel

 Numeric variables

– Policy value distributions

 Correlations

 Other issues

– Missing entire files

– Files with misleading names (e.g. wrong valuation date) – Inconsistent feature

descriptions (e.g. reset/ratchet) – “unreliable” fields

Takeaway: every variable that is part of a typical

experience analysis has required specific attention to ensure consistency and model validity, as have each of the additional 50+ fields.


Thank you!

Eileen S. Burns, FSA, MAAA Milliman

eileen.burns at milliman.com 206-504-5955


Predictive Modeling Case Studies

Liz Olson, FSA, MAAA



• I am not:

– A statistician

– A model builder

– Responsible for experience studies

• I am:

– An actuary

– Responsible for assumption governance

– Someone who knew there was power in statistics, but spent years

before finding the right statistician to pull it off


Experience Studies Where We’ve Used Predictive Modeling

• Variable Deferred Annuity Lapses

• Deferred Annuity Mortality

 Fixed Deferred Annuity Lapses

 VUL Life Insurance Mortality

 Life Insurance Renewal Premiums


Fixed Deferred Annuity Lapses

• Examined Fixed Deferred Annuity lapses using predictive modeling for

the first time in 2013

• Some factors that lapse behavior varied by were:

– Attained Age

– Policy Size

– Effective Surrender Charge

– Difference between credited rate and market rate

– Qual vs. Non-Qual

– Guarantee Minimum Floor Rate

• These drivers also had different sensitivities depending on whether the

policy was in the surrender charge period, in the shock year, or in the post-shock years


Fixed Deferred Annuity Lapses

• Typical graphs of dynamic parameters

• Predictive modeling allows for better delineation between base and dynamic lapses

Dyn am ic  Fa ct o r Policy Size

In CDSC Shock Post Shock

Smaller Larger Dyn

am ic  Fa ct o r Attained Age

In CDSC Shock Post Shock Younger Older


Fixed Deferred Annuity Lapses

• Progression of Models – Base Lapses Only

5 La p se  Ra te

Actual Lapse Rates vs. Model Over Time

Actual Model: Base Lapses Only


Fixed Deferred Annuity Lapses

• Progression of Models – Base and Some Dynamic Parameters

6 La p se  Ra te

Actual Lapse Rates vs. Model Over Time

Actual Model: Base + Some Parameters


Fixed Deferred Annuity Lapses

• Progression of Models – Final Model

7 La p se  Ra te

Actual Lapse Rates vs. Model Over Time

Actual Model: Final


Fixed Deferred Annuity Lapses

Lessons Learned

– Just because the model fits the past doesn’t mean it’s good for predicting

the future

– The model told us that for one product the difference between our credited

rate and the market rate didn’t matter for policy durations post shock

• The study had a few years of experience, but all in low rate environment

– Don’t rely on the model but rather use it to inform decision making

– Involve actuarial modelers early on to ensure the dynamic parameters are


– “Language barrier” between statisticians and actuaries


VUL Life Insurance Mortality

• We worked with a consulting firm to analyze our VUL mortality

experience using predictive modeling

• In addition to analyzing mortality across the obvious dimensions (age,

gender, underwriting category, etc.) we were also able to gather data from our P&C business or from other sources

– Home size, number of people in household, automobiles

– Interests of individuals in the household (i.e. boating, skiing, scuba, golf, organic foods, interior design, video games…)

• Consultant was able to tease out two profiles with very different

mortality experience

– Profile A: More attributes lead to lower mortality

– Profile B: More attributes lead to higher mortality


VUL Life Insurance Mortality

Customer Profile A Customer Profile B

0 0.5 1 1.5 0 1 2 3 4+ M o rt alit R e la ti vi ty Number of Attributes of Profile A Relative Mortality by Number of  Profile A Attributes 0 0.5 1 1.5 2 2.5 0 1 2 3 4+ Mo rt al it Re la ti vi ty Number of Attributes of Profile B Relative Mortality by Number of  Profile B Attributes 10


Life Insurance Renewal Premiums

• New project to use predictive modeling to establish renewal premium

assumptions for ULSG and VUL products

• Could ultimately be used as a retention strategy to proactively notify

customers who the model suggests may have a higher likelihood of stopping or reducing their premiums


Life Insurance Renewal Premiums

• Some companies may be setting this assumption in a very simplistic

manner (i.e. all policies make renewal premiums such that the aggregate matches expectations)

• This can lead to a sizeable error on many policies

12 Histogram of Actual vs. Predicted  Renewal Premium Error All PHs Pay Average No Error Large Error Large Error Po lic y  C o unt


Life Insurance Renewal Premiums

• Our new approach using predictive modeling has further improved our

renewal premium assumption

13 Histogram of Actual vs. Predicted  Renewal Premium Error Predictive Modeling Bucketing Algorithm All PHs Pay Average No Error Large Error Large Error Po lic y  C o unt


Other Possible Areas for Predictive Model Use

• Living Benefit Utilization

• Transfers between Fixed and Variable investments

• Life Insurance Surrenders

• Life Underwriting Decisions

• 401k, 457, 403b Case Lapses and Participant Deferrals

• Sales Distribution Tendencies

– Could possibly identify producer abuse such as STOLI or STOVA


Data Quality

• ASOP 23 provides insights into what an actuary should do when using


– Review data for reasonableness and consistency

– Not required to audit the data

• Some steps that we’ve taken

– Compare to ledger sources (policy count, deaths, lapses, FY

premium, renewal premium, claims)

– Cursory review of distributions (issue age, account value, face

amount, gender, underwriting category)


Staffing Model

• Pure statisticians often don’t have the product background or

business intuition needed

• Actuaries often don’t have the statistical horsepower needed

• Finding someone with both of these is rare

• Leaders could staff by:

– Using pure statisticians to build the predictive models and

using actuaries to manage the projects and add insights based on their intuition and experiences, or

– Grow the statistical talent by sending actuaries through

extensive statistical training so they can run the predictive modeling project from start to finish



• Seemingly endless possible uses of predictive modeling in

experience studies

• Think beyond just using traditional internal data

• Garbage In = Garbage Out

• Build the right team


Guillaume Briere-Giroux, FSA, MAAA, CFA

Integrating Predictive Analytics in Assumption Setting

Implementation and Integration in Financial Models 2014 SOA Annual Meeting & Exhibit



I. How and where predictive analytics impact assumption setting?

II. Implications for assumption setting process

III. Challenges and solutions for financial modeling integration IV. Key takeaways


How predictive analytics impact assumption setting?

B u si n ess v a lue

Data analytics literacy

Describe / monitor Analyze / understand Score / predict Decide / optimize / manage Descriptive analytics

(what happened and why?)

Predictive analytics

(what will happen?)

Prescriptive analytics

(what should we do?)

Scope of predictive modeling techniques Enhanced experience studies Enhanced assumption setting Enhanced model-based decisions


Where predictive analytics impact assumption setting?

Use of predictive modeling is increasingly widespread for experience studies

Product Surrenders / Lapses

Utilization / funding

pattern Mortality Morbidity

VA Living Benefits

FIA Living Benefits

Fixed Annuities

Universal Life


Long Term Care

We are also seeing greater use of predictive analytics in M&A


Implications for assumption setting process


More attention is paid to more secondary internal variables


Increased opportunities to test external variables


Additional relationships to study and understand


More comprehensive ”data-driven” discussions


Better ability to put experience in context

In summary, using predictive analytics requires more resources dedicated to assumption setting but enables richer thinking around key assumptions


Example: Integrated GLWB policyholder behavior cohorts

Cohort of GLWB Observed behavior

“Efficient” users • Utilize 100% of GLWB maximum income • Strong utilization “feature skew”

• Low lapse rate

• More efficient dynamic lapses

“Partial” users • Utilize less than 100% of GLWB maximum income • Weaker utilization skew

• Higher lapse rate than efficient users • Less efficient dynamic lapses

“Excess” users • Utilize more than 100% of GLWB maximum income • Very high lapse rates

• Least efficient dynamic lapses

“Waiting” users • Have not yet utilized • Low lapse rates

• Efficient dynamic lapses • Waiting for rollup?


Computational implications for modeling

Run time optimization becomes a three dimensional problem

Can specify accuracy functions to determine optimal accuracy for a given run time and generate an efficient frontier


Other considerations for modeling

How granular should the model become?


Materiality and certainty of dynamic


Materiality of business


Model purpose


Degree of buy in


Model implementation approach

There is a compromise between transparency, flexibility, controls and system performance

Desirable property Parameterized formula Factor tables


Flexibility (model form)

Flexibility (adjustments) Ease of control Auditability Computational performance High Low


Best practices for model implementation

Area of focus Considerations

Parallel testing • New assumptions are more complex to code • Do single cell-testing with replicator (e.g., Excel)

• Excel replicator can also be used for extreme value testing / sensitivity testing

Internal data • Data definition between experience study and financial models must be consistent

External variables • Modeler must understand the sensitivity of projected behavior to external variables (how reliable are my scenarios?)

Sensitivity testing • Understand the potential impact from stress testing the assumption parameters

Documentation • Documentation of rationale for key modeling decisions and any limitations or simplifications


Key takeaways


Think about the business and the environments


Think about the models and their end goal


Related documents

The assumption of uniform impatience has important implications for asset pricing as it rules out speculation in assets in positive net supply for deflator processes in the

In children, a prolonged cough, usually following a viral upper respiratory tract infection is often diagnosed.. any benefit of

Five feature detectors were tested based on their quality of repeatability, using two faces from the same object, which were taken from left and right, in order to simulate

Concentration of daily rainfall obtained through the CI confirmed that south and central Chile have similar values as those of the Iberian Peninsula, high values of CI for

Box-plot for the target position error obtained by each one of the five proposed methods for the beta band Source: Created by the authors.. Distance to target in mm achieved by

The companies of which the financial information has been included in the Issuer’s reviewed Interim Condensed Consolidated Financial Statements for the six-months ended 30 June

suggest that the accuracy of these judgments was not simply a function of individuals accessing and using their perceptions of relationship quality to assess the likelihood of

Following the argument of Cohen, Levinthal then, it is firms with higher absorptive capacities in a cluster that are more likely to establish linkages with external sources