Session 058PD: Artificial Intelligence for Actuaries
10/15/2018
3:30-4:45 p.m.
SOA Antitrust Compliance Guidelines SOA Presentation Disclaimer
http://www.soa.org/legal/antitrust-disclaimer/https://www.soa.org/legal/presentation-disclaimer/
Artificial Intelligence for ActuariesModerator: Sarah Abigail
Presenters: Shankar Vaidyanathan
Martin Snow, FSA, MAAA
Gaurav Gupta
October 15, 2018
#AIforActuaries
2
Sarah Abigail
Co-Founder Ironbound Consulting Group
@ironbcg
Shankar Vaidyanathan
Founder and CEONoonum
@_noonum_
Martin Snow, FSA, MAAA
VP, Chief Delivery OfficerAtidot
@AtidotIsrael
Gaurav Gupta
Founder and CEOQuaEra Insights
@QuaEraInsights
Artificial Intelligence for Actuaries
3
4
Artificial Intelligence & Machine Learning PAST, PRESENT, AND FUTURE
Shankar VaidyanathanFounder & CEO, [email protected]://linkedin.com/company/noonum@_noonum_
http://linkedin.com/company/noonum
5
What is Artificial Intelligence?Human intelligence "can be so precisely described that a machine can be made to simulate it".
6
Intelligence
• Learning – rote learning and generalization• Reasoning – inductive and deductive• Problem solving – special purpose or general
purpose
• Language – Text, Speech, Images, Video• Perception
7
Origin
• Alan Turing• Idea of machine intelligence• Cryptanalyst during World War II
• Workshop at Dartmouth College in 1956• Allen Newell (CMU)• Herbert Simon (CMU)• John McCarthy (MIT)• Marvin Minsky (MIT)• Arthur Samuel (IBM)
8
Chess
• Drosophila of AI• Methods of problem solving and learning tested• Use heuristics to narrow down moves• 1997 – computer beat the world champion• Still a lot of computing involved
9
Self Driving Car
• 2 million miles of driving experience• Constantly processing data from
sensors
• Learn to respond to risky situations effectively
10
Companyof
Interest
Customers
Competitors & Disruptors
Fundamentals
Economic
Sentiment
Suppliers & Partners
understands companies
11
Market (Stock/Bond)
CompanyFundamentals
Economic
Social & News Media
Analysts reports
Alternative
Other (Weather,
Health)
BYOD
insights
is constantly learning…
12
Other Applications
Finance
InvestmentsCorporate Strategy
BankingActuarial
CryptocurrenciesBlockchain
Health Care
Hospital readmissionPatient Care
Improvement
Industrial
Internet of ThingsFailure Prediction
Consumer
Smart PhonesIntelligent Homes (IoT)
…
13
Growth in AI and Machine Learning
Cloud• Powerful Computing Machines• Availability of lots of Data and storage capacity• Connectivity
Public libraries
14
Future Challenges
• Ethical decisions • Drawing inferences relevant to a
situation
• Connotation and words with double meanings
• Non-verbal cues• Emotions in language
15
Future Advances
Improvement in Storage and Processing
Quantum Computing
Improvement in Connectivity
Better Implementations
16
Thank You
Shankar VaidyanathanFounder & CEO, [email protected]://linkedin.com/company/noonum@_noonum_
http://linkedin.com/company/noonum
Emerging Uses of Predictive Analytics
SOA 2018 Annual MeetingSession 58
Martin SnowVice President & Chief Delivery Officer
Member Advisory BoardAtidot
October 15, 2018
2
Predictive Analytics Data Flow
Additional TablesOne view of the client
Policy Administration Transactions
Customer RelationshipManagement
Business question
Historical / Snapshot data
Feature engineering
External data
The modeling Arena
Results & Insights Actions
Validation
Business Questions and Modeling
3
Use Case: Retention
Business Question: Can we determine who is likely to lapse and why to improve retention?
Model Objective: Predict who is at risk of lapsing to foster conservation efforts.
Model Considerations:May want to fit several models – based on product complexity, variation by policy duration, data availability etc.
What the Model Does: Classifies policyholders based on propensity to lapse using supervised learning. Validation: Based on observed past experience.
Business Questions and Modeling
4
Use Case: SalesBusiness Question: Is there an underinsurance opportunity?
Model Objective:Find outliers with less insurance than peers; Rank them based on propensity to buy (and need for) more insurance; Provide insights on the influencing features
Model Considerations:Unsupervised clustering of similar individuals vs. supervised regression over the amount insured; width of distribution around the mean; model overfitting
What the Model Does:Group policyholders based on similar characteristics –policy type, insurance characteristics, demographic information – using regression analysis.
Validation: ?
Feature engineering
5
Feature engineering is the process of using domain knowledge of the data to create features that make machine learning algorithms work.
Feature engineering is fundamental to the application of machine learning, and is both difficult and expensive.
(Wikipedia)
Feature engineering - examples
6
How would we incorporate the issue date of 11/22/2017 ?Feature Meaning
Date Models long term trends
November Month-specific sales features, year end targets
22nd day of the month Month end targets?
Wednesday Significance of day of the week
Proximity to public holidays Policyholder behavior
Proximity to financial / political events ….. Policyholder behavior
Premium / Contribution information: Missing / Skipped premium - How many times did a policyholder miss a premium?
- How recently?
- How to combine the information – for example:Missing Premium = (number of times missed in last 24 months) * (1 / distance of last one)
Feature engineering - examples
7
“Free-Text” professions:
- Thousands of different occupations, not useful for analysis
- Used advanced clustering techniques to map to 10 groups
- Result: Occupation is significant for propensity to lapse
- Another enhancement – connect to underwriting risk classification
External Data
• Most common – Demographic data based on address, for example: Median earnings, average household, homeowner vacancy, median age
• Lifestyle type data (subscriptions, etc.)
• Depending on product, market data can also be useful
• Life cycle events, from external sources, can be useful triggers, for example:
• house purchase
• Job change
• New family member
• Company data which is “external” to the block analyzed – for example from other operations (health, P&C?)
8
All data and (engineered) features can be useful and may lead to powerful insights –
Time spent here is usually well rewarded!
Validation (lapse example)
• How do we know the model is “working”?
• Back-testing against historical data and against company assumptions:
• Define “training” period (eg 2010-2015)
• Defined “back testing” period (eg 2016)
• Check actual and own company assumptions against model results for 2016
• Continuous monitoring throughout
9
Insights – Example - Persistency by features
10
Pensioner Clerk Housewife Teacher Manager
• Once model is validated, we get three useful outcomes:
• Feature importance
• Predictions on a per policy level
• Ability to predict target based on simulated input
• Features can be grouped into two categories:
• Indirect (e.g., occupation)
• Direct (i.e., company can influence)
• External (e.g., market interest rates)
Insights – Example - Persistency by features
11
Insights into actions and simulations
• For a lapse / underinsurance model, results are probabilities per policy. The model can assist in prioritizing policies for conservation or up-sell:
• Either when the client approaches the company
• Or proactively
• By adding an additional layer of per-policy profitability, insights turn from inforce projections to value projections
• In addition, model can be used to simulate business under various scenarios.
• Optimization techniques can be applied to select the preferred outcome
12
Fixed Annuities and Reinsurance
• Low interest rate environment leads to spread compression and lower profit margins
• Companies are not prepared to invest more capital
• De-risk the balance sheet, diversify the business and improve investment margins
• How well does the cedant understand its lapse experience?
• How well does the reinsurer understand its lapse experience?
13
Fixed Annuities Lapse Experience
• What data do you have?
• Do you have the skills to use machine learning and predictive analytics on this data?
• Do you have data scientists with expertise in the annuity business?
• What type of data have your machines been trained on?
• What type of turn around do you want from predictive analytics?
14
Some closing thoughts
• Predictive analytics can be a powerful tool in managing in force business and generating new sales
• Results depend on availability of data. Frequency of client contact impact results. Use of external event data can substitute for missing internal data, but is more difficult / expensive to obtain.
• Feature engineering and augmentation are critical.
• We have not discussed feature correlation and masking but these are important issues which are tricky to handle
• Additional layers (e.g., profit) can be incorporated for simulations to help with strategic decisions
• Best results are achieved when predictive analytics are integrated to the business process
15
Artificial Intelligence for ActuariesModerator: Sarah Abigail
Presenters: Shankar Vaidyanathan
Martin Snow, FSA, MAAA
Gaurav Gupta
October 15, 2018
#AIforActuaries
2
Sarah Abigail
Co-Founder Ironbound Consulting Group
@ironbcg
Shankar Vaidyanathan
Founder and CEONoonum
@_noonum_
Martin Snow, FSA, MAAA
VP, Chief Delivery OfficerAtidot
@AtidotIsrael
Gaurav Gupta
Founder and CEOQuaEra Insights
@QuaEraInsights
Artificial Intelligence for Actuaries
3
Artificial Intelligence for Actuaries
Session 058PD
SOA Annual Meeting, 2018
Gaurav GuptaFounder & CEO
How Can YOU Use it?
8
Powering Life Insurance with AI techniques
Supervised learning
Deep learning
Reinforcement learning
NLP
Bayesian
Regression
ClusteringDimensionality
reductionSVM
GAMRNN
Simulation
Recommendation systemRandom forest
Decision trees
GLM
Monte Carlo
AI
How to build the bridge?
Retention
Life Insurance
Cross-sell
Pricing
Reserving
Product development
Anti-selection
Distribution
Assumption setting
Mortality study
Reinsurance
Gradient boosting machine
Emerging experience
Risk management
Premium persistency
Underwriting
9
PRICING | UNDERWRITING | RETENTIONPRODUCT DEVELOPMENT | RESERVING
CROSS-SELL | PREMIUM PERSISTENCYMORTALITY STUDY | ANTI-SELECTION
RISK MANAGEMENT | ASSUMPTION SETTING
Where is AI being used in Life Insurance?
AI + P&C
RETENTION
PRICING
UNDERWRITING
PRODUCTDEVELOPMENT
LAPSE
RESERVING
CROSS-SELL
PERSISTENCY
MORTALITY STUDYANTI-SELECTION
EXPERIENCE STUDYASSUMPTION SETTING
AI + LIFE INSURANCE
Where is AI being used in Life Insurance?
* Data collected from Google scholar search, number of articles10
RISK MANAGEMENT
11
What if you could…
… Monitor mortality constantly
… Add and analyze unlimited data, new variables quickly
… Predict policyholder behavior (shock lapse, anti-selection, …)
… Develop new mortality assumption in half an hour
… Price at an individual level
… Develop customized products on-the-fly
12
Why AI has not been fully exploited by life actuaries
• Hurdles for AI to fit existing paradigm
• Talent to understand both AI and life insurance is scarce
• Difficulties in changing business processes and systems
• Lengthy validation
• Death is a rare event, and has a long duration – low credibility
• Inconsistency of data dimensions, variable definitions etc. across systems, over time
• Blackbox is a no-no• Regulation
Regulatory Data Other…
Case Study:Experience Study
Identify segments of mortality deviation Integrate underwriting, transaction, claim, external data in real-time Discover key drivers for mortality/lapse deviation Detect early alerts for better/worse mortality trends Generate reports in minutes instead of weeks
AI can help you do things faster and easier
Example Areas Experience studies Automate reporting
How AI Achieves it More powerful data handling - 1) more data sources; 2) easier data manipulation; 3) faster Easier to develop and monitor KPIs (e.g. deviations, trends) Large selection of algorithms, software and platforms
Identify segments of mortality deviation
14
~100%
14
Actuals to Expected (A/E) ~500,000 exposures
>120%
Case Study:Assumption Setting
Develop better assumptions that satisfy inherent dimensional relationships Develop assumptions by channel of acquisition Provide more accurate forecasting for future events (death, lapse, surrender, etc.) Estimate marginal impacts of underwriting requirements
15
AI can help you do things better
Example Areas
Assumption setting Projection of claims, lapses, surrenders,
withdrawals Protective Value studies
Segmentation Reinsurance structuring and pricing Stress and scenario testing
How AI Achieves it Supplement actuarial credibility with AI validation assessment metrics Automatically detect and utilize correlations and interactions Tools to capture more reliable, sustainable relationships to minimize overfitting
Develop better assumptions that satisfy inherent dimensional relationships
16
Use case - AI can help you do things better
Female, NS* VBT 2015 table
AI can …• Borrow information from “adjacent”
cells along multiple dimensions• Enforce complex relationships
• Select, ultimate periods• Omega rate• Monotonic constraints • …
Develop better assumptions that satisfy inherent dimensional relationships
Duration
Issu
e Ag
e
The cell does not have enough credibility to make any adjustment
With borrowed information from “adjacent” cells, we can be more confident to make adjustment
AI can help you do new things
17
Example Areas
Completely fluid-less underwriting New underwriting, pricing process powered by image, text, voice, cognitive AI New capability to provide more granular (closer to individual level) pricing Personalized product
How AI Achieves it Capability to handle large volumes of data Create more powerful features via combinations of image/text/voice/cognitive AI Larger selection of algorithms
Future Use Case:“On-demand Insurance”
Create personalized life products – variable duration, structured riders, unique risks, etc. Determine individualized pricing for the personalized product Determine, gather and process all sources of data on the applicant, coverage and scope
(e.g. IoT, images and voices, social media, circumstance specific risk factors, etc.) Estimate mortality risk for that circumstance (e.g. traveling to foreign countries,
participating in risky events, covering for predefined period, etc.)
18
How to get started• Pick someplace to start
• Small enough problem to finish in 3-5 months• Delivers business value
• Spend time defining the business problem • Example:
• Do I want to predict mortality or mimic the underwriter decision in my predictive underwriting model?
• Do I want to optimize response to direct marketing campaigns or response of prospects with the best risk profile?
• Other?• Start with the data you have
• The best source of data is your own. You can more out of it than you think• More data does equal better results…harder to find a needle if the haystack is bigger.• There is always more thing you can find from your current data
• Assemble the right team• Actuaries + Domain Experts + Data Scientist
19
Thank You!
If you have more questions on AI, please come see us at Booth 519 in the Exhibit Hall, or contact me at [email protected]
Cover pageVaidyanathanSnowGupta