iowa actuaries club education day - big data and forward ... · predictive modeling relies on...
TRANSCRIPT
Iowa Actuaries ClubEducation Day
Big Data and Forward LookingModels – a life perspective
February 25, 2014
Purpose and topics
2
In this shortpresentation we’llexplore where the lifeinsurance industrystands today, what bigdata and forwardlooking models mean tothe industry, and wherechallenges lie.
BehavioralSimulations
ConclusionIntroductionAnalytics
trends
Current state Challenges
Policyholder behavior survey
3
PwC, together withLIMRA, conducted asurvey of how companiesare modeling policyholderbehavior and settingassumptions.
In this section we discusssome of the key findingsfrom the survey and ourresearch.
Behavioralsimulations
ConclusionIntroductionAnalytics
trends
Current state Challenges
Survey Design
4
Reviewed over 100 articles, papers and books in the following generalareas:
1. Academic2. Actuarial3. Industry
An electronic survey with over fifty questions was sent to over 100 life &annuity companies.
Conducted in-depth qualitative interviews with:1. 8 Life insurance companies2. 3 P&C insurance companies3. 3 Non-insurance companies
Literature Review
1
QuantitativeQuestionnaire
2
QualitativeInterviews
3
Key Findings
1. The life insurance industry is behind the P&C insurance industry and other industriesin using advanced analytical techniques to understand their customers (i.e.,policyholders).
5
“Information Advantage” Development StagesBuilding an “IA organization” is a multi-phase process that requires sequentialinvestments to develop capabilities over time
• Integrated internal andexternal data to createnew solutions
• Fully simulated businessoperations toevaluate decisions
• Pervasive institution ofa Service OrientedArchitecture
• Increased enterpriseinformation security
• Analytics provideenterprise wide insightsin real time
• Advanced modelingbeing used to evaluateacross functional areas
• Information integratedacross the organizationfrom a single,recognized source
• Increased role and userbased security
• Some modeling abilitywithin functional areas
• Functional areas havedeveloped high dataquality within their areas
• New measurementpoints added tooperations processes
• Canned reports withlimited usability
• Little-to-no modelingfunctionality
• Data quality reliantupon significantmanual adjustmentsT
ech
nolo
gy
• Delayed access to data
• Functional integration,aggregation
• Limited and delayedmetrics
Org
aniz
atio
n • Multiple manualprocesses to enter,correct, extract,manipulate andanalyze data
• Specialized reportspulling from multiplesourcesare available uponspecial requests
• Business users configurereports dynamically withnear real time information
• New tools and traininginstituted to betterextract business valuefrom information
• Improved culture ofinnovative thinking thatidentifies new info.opportunities to test,learn, and scale
Deci
sion
Maki
ng
• Delayed decision-makingbased upon limitedinformation and insights
• Improved insightsthrough special projects,butnot scalable
• Improved insights,decision-making andspeed of decision-makingwithin silos
• Enterprise-wide decision-making with ability tomake informed trade-offs
• Consistent fact-base foruse across theorganization inoperations, reporting anddecision-making
Level 1
Limited
Level 2
Evolving
Level 3
Functional Excellence
Level 4
Integrated Excellence
Level 5
Information Premium
Life InsuranceIndustry
P&C InsuranceIndustry
ConsumerProducts Industry
6
What are the challenges?
A number of limitations were often cited:
• Companies have put a lot of reliance on historical averages, and many failed to anticipate howpolicyholders behave under different circumstances;
• At the same time, data credibility is a limiting factor – and is particularly challenging whenthere is such a reliance on historical averages;
• Actuarial resources are limited – we are increasingly being asked to do more with less;
• Nonetheless, many companies recognize the shortcomings of traditional actuarial methods, andmany have at least considered implementing predictive analytics.
7
Upwards of 40% of insurance companies are using or considering using predictive modeling tobetter understand policyholder behavior1
How can we build better forward-looking models?
1. “Report of the SOA Predictive Modeling Survey Subcommittee.” Society of Actuaries. (Jan, 2012)
Analytics trends and the life insurance industry
8
In this section we’ll lookat some of the Big Dataand analytics trendsgoing on, and inparticular we’ll explorehow they apply to thelife insurance industryand where thechallenges are.
BehavioralSimulations
ConclusionIntroductionAnalytics
trends
Current state Challenges
Reduced Data Storage Costs
Over the last 30 yrs storage spaceper unit cost has doubled ~ every
14 months (increasing by an orderof magnitude every 48 months)
Big Data, Smart Analytics, and Sensor Devices are changing theway industries harness information
9
Analytic Techniques and Data Sources – Increasing Sophistication
Source: Property & Casualty Predictive Modeling Survey — Results and Implications." Towers Perrin Predictive Modeling, PwC Analysis
80’s-90’s
Multivariate Analysis
00’s-10’s
90’s-00’s
Data driven predictivemodeling
Co
mp
uta
tio
na
lP
ow
er
Ty
pes
of
Da
ta
Interpretation, Simulation& Visualization
Structured Data andSolutions
UnstructuredData andInsights
More Analytics Power
Computing power project tosurpass power of human brain by
2023
More Applications
Sequencing costs have dropped bya factor of 14,000 over the pastdecade, ~ 100 times faster thanMoore's Law in semiconductors
More Data Generated
“Every Six Hours, the NSA Gathersas Much Data as Is Stored in the
Entire Library of Congress or ~10TBs”
Acceleration Laws Apply to Analytics & Data
Why has the industry been slower toembrace big data and even predictiveanalytics?
10
Predictive Modeling Relies on Historical Correlations
11
StatisticalModel onHistorical
Data
Age
Credit score
Marital Status
Lapse Rates= f(age, credit score, marital status)
Predictive modeling more effectivelyaccounts for multiple variables comparedto traditional modeling. This isparticularly useful in testing forinteraction effects between variables (forexample, the interaction between theimpact of age and income on lapsation).
In some situations, this can help a lot withdata credibility issues.
However, predictive modeling relieson historical correlations to predictfuture results, without reflecting theactual drivers behind apolicyholder's decisions.
Overview of Predictive Modeling(illustrative)
Predictive Modeling Relies on Historical Correlations
1. Eric Siegel, Predictive Analytics: The Power to Predict Who Will Click, Buy, Lie or Die, Wiley (2013).12
• Following the prior example, the input factors such as the individuals age, martial status andcredit score don’t change even as the world around the individual changes; therefore thepredictive score for the lapse rate doesn’t change, either.
• The predictive power of statistical models significantly decreases when there is a broad shift inthe environment, which will generally happen when you are looking over longer time horizons(3-5 years).
• Being able to re-run the model frequently in the future does not obviate an insurers needs forstrategic planning and decisions made today.
• This does not mean traditional actuarial methods are better –they suffer the same drawbacks.The only advantage is that they are slightly more intuitive/simple, and this transparency canhelp you understand the limitations of the modeling when you have a complex problem.
“[Predictive modeling] is designed to rank individuals by their relative risk, but notto adjust the absolute measurement of risk when a broad shift in the economicenvironment is nigh.” 1
Behavioral Simulations
13
In this section wediscuss how a morerobust forward lookingmodel can be built.
These modelingtechniques have onlyrecently become possiblefor the life insuranceand annuity area andare an excitingdevelopment.
BehavioralSimulations
ConclusionIntroductionAnalytics
trends
Current state Challenges
Literature Review: Agent-Based Modeling
14
Agents of Change, The Economist, July 22, 2010
The assumptions, including efficient financial markets and rationaleexpectations, are considered to be too simplistic.
A new approach called agent-based modeling is being explored to helpaddress lessons learned from the financial crisis.
Mills, Alan. Complexity Science: An introduction (and invitation) foractuaries, Society of Actuaries (2012) Understands the complex nature of social systems. To grasp and manage the systems in which actuaries work, we must
augment our tools with the new methods of Complexity Science.
Mills, Alan. Simulating health behavior: A guide to solving complex healthsystem problems with agent-based simulation modeling Society ofActuaries (2013) Agent-based modeling simulates agents’ (e.g., individuals and
companies) interactions with their environment and other agents The goal is to understand the emergent behavior of complex systems.
1
2
3
Understanding the policyholder
15
It is important that we view thepolicyholder not as a male age 40nonsmoker, but as the member ofsociety and part of a householdwith a focus on:
• The composition of thathousehold and how it changesover time;
• The life events that take place inthe household such as havingchildren;
• The household’s income,spending, and savings habits;
• The type of assets the householdowns and the liabilities thehousehold owes; and
• The choices the household makes,both rational and behavioral.
Life Events
• Getting married
• Buying a house
• Having a child
• Retiring
Income Statement
• Salary
• Expenses
1. Nondiscretionary
2. Discretionary
3. Health costs
Balance Sheet
• Assets
1. Home
2. Financial assets
• Liabilities
1. Mortgage
2. Personal debt
Choices
• Rational
• Behavioral
1. Mental accounting
2. Joint decisionmaking
3. Financial literacy
Understanding life events and choices
16
The goal is to understand howlife events and the choices anindividual make change overtime such as when:1. He or she graduates from
college and gets a job;2. He or she marries;3. They have children;4. They become “empty
nesters”; and5. They retire.
Dependents Single & ‘Rich’ Growing Family Pre-Retiree Retiree New Generation
Liability Creation
Asset Transfer
Asset Creation Asset Creation
Asset Protection
Asset Preservation
Asset DepletionP
olic
yho
lde
rL
ife
-Cyc
leS
tag
es
Life
Eve
nts
Ad
vic
e
Asset Cycle
• Paying off student loans
• Starting a career
• Getting married
• Paying off student loans
• Starting a career
• Getting married
• Buying a home
• Having or adopting children
• Buying a home
• Having or adopting children
• Paying tuition bills
• Caring for parents
• Planning for retirement
• Paying tuition bills
• Caring for parents
• Planning for retirement
• Withdrawal money for retirement
• Paying for health care
• Creating a legacy
• Withdrawal money for retirement
• Paying for health care
• Creating a legacy
Behavioral Simulation
Simulation of how individualsreally make decisions and theiremergent group behaviorsbased on modeling individualbehaviors as ‘agents’. Choicemade by individuals getreflected as ‘market-level’emergent behaviors that arecalibrated with actual andsurvey data
Behavioral simulation combines agent-based modeling andbehavioral economics to model individual decision-making andemergent behaviors
Artificial IntelligenceCognitive thought throughmachines
Complex SystemsEmergent systembehavior from individualactions
Computational PowerRapid cycle-timefor intensive calculations
Agent Based Modeling
Sophisticated, computationallyintensive modeling techniquethat relies upon a decentralizedset of behavioral rules andstudies emergent behaviors
Classical EconomicsIndividual decision-makingdriven by self-interest andutility maximization
PsychologyScientific study of mentalfunctions and behaviors ofindividuals and groups
Behavioral Economics
Study of individual decision-making based on cognitive,heuristic, emotional and socialfactors
+
+
+
+
=
=
=
17
Enabling factors:
1. Big data – significantly increaseddata available to study consumerand policyholder behaviors.
2. Predictive analytics andmachine learning are needed toidentify drivers of behaviors and tocalibrate agent decision rules.
3. Increased computing power tosimulate millions of consumerbehaviors.
4.Developments of software anduser communities for agent-based modeling beyond pureresearch.
Example of annuity withdrawal and surrender drivers
18
1 2 3
4 5 6
Thank you!
19