AIandAdvancedAnalytics:AugmentedUnderwriting
LeeSarkinHeadDataAnalyticsL&HAPACMEA
MunichRe,Singapore
1. Munich Re: an integrated analytics reinsurance partner
2. Augmented underwriting
3. Replacing traditional underwriting with external data
Agenda
Munich Re: an integrated analytics reinsurance partner
1
Munich Re: an integrated analytics reinsurance partner
Life clients • 5 in Asia
>200 >100 >40
Data scientists
Analytics pilots
Global footprint Allows local learnings to be scaled
Proven competence Portfolio of real cases applying insurance domain and analytics expertise to meet insurers’ needs
Research & development Significant global R&D in AI and analytics for insurance
Image: used under license from shutterstock.com
Banks • Amongst the largest in Africa • Amongst the largest in Asia
Asian retailer • Amongst the largest in Asia
and top 30 globally
We accept the resulting insurance risk Alignment of interest with clients on risks emerging from streamlining underwriting
Integration of auto-uw and predictive modelling Predictive models can be deployed at PoS via MRAS to drive digital predictive underwriting offers
Hong Kong
Singapore
Japan
Indonesia
Vietnam
Thailand
Dubai (Middle East)
Beijing
Malaysia
W
Image: used under license from shutterstock.com
The four pillars of data analytics
… Creating data-driven business value requires an integrated proposition…
Integrated analytics: § Internal § External § Structured § Unstructured
Data
§ Hardware § Software
Technology
§ Data scientists &
engineers § Actuaries § Business people § IT architects
People
§ Regression
models § Machine Learning
models § Text mining
Methods
Gains stakeholder buy in
Implementable
Treats customers fairly
Risk acceptance
Auto UW
Risk Selection
Eligible Customers
Digital Sales
Distribution Method
Product Complexity
Data analytics
Pricing
Full UW
Risk Appetite
No Traditional UW
Reinsurance expertise
Simplified underwriting with minimal price increase
Image: used under license from shutterstock.com
Why an integrated approach?
W
Conversion of analogue data into digital form that can be used by a computer
Digitisation
Manual to automated processes Automation
Data management Data Capture
Analysis
Digital transformation
Digital transformation
Prescriptive analytics
Predictive analytics
Diagnostic analytics
Descriptive analytics
What should happen?
What will happen?
Why did it happen?
What happened?
Gather data
Perform analytics
Risk assessment Risk categorisation
Targeted underwriting based on insights and risk assessment
Customised products based on customer insights
Model deployment, customised campaign design & execution
Underwriting & product
Deploy model
Campaign design & execution
Monitoring
How to get the balance right?
Munich Re accepts the resulting insurance risk
Analytics Proposition development Risk acceptance
Integrated analytics: Data-driven business value in four integrated steps
Image: used under license from shutterstock.com
E.g. simplified underwriting, products, pricing, risk appetite?
How to get the balance right?
Define the problem
Data
s s
s
Build model Develop proposition
Deploy model Campaign design and execution
Monitoring
…We accept the resulting insurance risk
Pricing and product development Underwriting Portfolio
management Claims Sales & marketing
Which underwriting questions could be removed with minimal or no price impact?
How to identify customers with greater likelihood to purchase?
How to identify which customers are eligible for products requiring less or no underwriting? How to eliminate
underwriting requirements using external scores in underwriting?
How to use bank or physical activity data in underwriting?
How can I predict which customer smoke?
How can I predict fraudulent claims?
How can I automate claim decisions?
How can I automate aspects of experience analysis when deriving best estimates?
How to develop a product based on steps per day?
Which customers are most likely to lapse?
Can predict the time fo return to work for disability claims in payment?
How to enhance the customer’s underwriting journey focusing on immediate underwriting decisions?
Access data from various sources
Perform analytics and gather insights
s s
s
Low risk Average risk High risk
Risk selection / categorisation
Insights
Image: used under license from shutterstock.com
Proposition development
Insights Lead generation
Standard rates with reduced underwriting and evidence requirements
Refer to underwriter
Increased STP rates
Improved take-up rates of cross and up-sell offers
Enhanced general issue (GIO) / simplified issue (SIO) product offers
Optimised underwriting rules via predictive analytics and deploy at an electronic point of sale
Reduced non-disclosure / anti-selective behaviour
Reduced lapses
Reduced claims fraud Automated claims decisions (increased claims STP)
Predict cause of claim from claims reports
Increased efficiency of experience analysis Increased accuracy of best estimates and reduced operational risk
Reduced underwriting with minimal / no increase in pricing
Reduced cost of underwriting without increasing risk
DOB Gender Smoker Income Marital Status
BMI UW Decision
7/24/1975 F N 240000 Married 30 S7/17/1969 F S 552000 Widow 22 NS7/18/1973 M N 180000 Married 30 S3/18/1971 F N 200000 Married 20 S2/27/1970 F S 352000 Married 20 S12/29/1975 F S 460000 Married 19 NS7/22/1986 M N 850000 Married 34 S7/24/1975 F N 240000 Married 30 S7/17/1969 F S 552000 Widow 22 NS7/18/1973 M N 180000 Married 30 S3/18/1971 F N 200000 Married 20 S2/27/1970 F S 352000 Married 20 S12/29/1975 F S 460000 Married 19 NS7/22/1986 M N 850000 Married 34 S
Historic records Known outcome
Build model
1
DOB 7/27/1975Gender FSmoker NIncome 240000Marital Status Married Real Weight 70
New customer record
P(Standard) 0.78
Deploy model
2
Predict outcome
3
Model
s s
s
Solution in the Cloud
Own Deployment & Integration System
Deployment options
Consume model
Model deployment, customised sales execution method
Outbound
Inbound
Hybrid
Preselect customers with offers through direct marketing channel
(lower risk)
Open to new business based on predictive variables with
offers through digital channel (higher risk)
New business profile and agent analysis
Random holdout sample
Claims experience
Trends to monitor
Risk acceptance
Example proposition outcomes
Across the whole value chain
Augmented underwriting powered by AI Asian insurers are streamlining underwriting with AI…trendsetter
1
Image: used under license from shutterstock.com
Augmented underwriting proposition
30%
67%
3% Pain point: Number of questions Pain point:
Low STP
Pain point: Are medical evidence
requirements cost effective?
99%
0.89%
0.01%
80%
19%
1%
Pain point: Free text boxes
Pain point: Are NMLs cost-
effective?
Pain point: Was a referral unnecessary?
Potential outcomes: Increased STP
Reduced questions with minimal / no price impact
Reduced cost of underwriting
Improved customer experience
Enhanced products based on customer analytics insights Increased cost-effectiveness of medical evidence and NMLs Improved future profitability or more competitive pricing
Final decision
Final decision
Case study 2
>60% reduction in the number of referred cases (72% of all cases are currently referred by AIS) means GE can redistribute this time to other activities and eliminate opportunity cost Increased capacity to underwrite significantly larger volumes without employing more underwriting (increasing cost) and operational risk >60% reduction in the number of referred cases (72% of all cases are currently referred by AIS) enables GE to redistribute this time to other activities and eliminate opportunity cost
Improved agent and customer experience from eliminating long reflexive rules that result in referrals Significant cost reduction for GE Yes
Objective
Double STP with minimal risk increase
Reduction in application form questions with minimal increase in risk
For referred cases, assess the value of medical evidence Enhance simplified underwriting propositions by identifying customer segments that qualify for no or reduced underwriting
Human Resources Q1 OKR
Doubled STP
Less referred cases (reduced manual uw and costs)
Reduction of medical evidence for straight-through standard cases
Reduction in base questions (“BQ”) for straight-through standard cases
>60%
100%
~50%
>100%
Key results Asian insurers use AI to streamline underwriting…trendsetter
Reduction in reflexive questions for straight-through standard cases 100%
Expected business value
Better experience for applicants
Better experience for distributors
Reduced manual underwriting
Reduced operational costs
Lower NTUs
Higher sales
Faster turn around time
Reduced availability of disclosures in claims assessment
Potential business impact
Data for AI model, risk evaluation and testing
§ Additional information gathered in later stages of underwriting
§ Examples: § Exclusions § Reasons for decline § Loadings on cases
§ Case data includes standard data components of each case
§ Examples: § Age & gender § BMI § Sum assured § Risk types § Internal Flags
§ This data reflects the disclosures made on the medical conditions from the base questions
§ Examples: § Epilepsy § Diabetes § High cholesterol § Cancer
§ Base question answers § Examples:
§ Heart conditions § Diabetic conditions § Family history
ANALYTICS BASE TABLE
§ Includes backend decision from underwriter § Notes/comments from underwriter
ü Model uses: underwriting data from an auto-underwriting engine and back-office systems
ü Prediction of final underwriting decision
Modelling the final underwriting decision – comparing machine learning models GBM performs best and incorporates hundreds of decision trees
§ Many machine learning models are compared to identify the best performing model
§ GBM, Xgboost incorporate hundreds of single decision trees and generally are more accurate than single decision trees
25% 75%
*Area under the curve (AUC) is a measure of predictive power. A model that randomly decides the underwriting decision would produce an AUC of 0.5 and a good model over 0.85
Mod
el ty
pe
AUC* Median
FOR ILLUSTRATION ONLY
Modelling the final underwriting decision (single decision tree) Which cases may have potential to be automated as ‘accepted standard’ (STP)?
Represents > 80% of cases within which >95% received a final standard decision
Sample application form Back office flag
BMI < X
BQ1
BQ2
BQ5
BQ7
BQ4
FOR ILLUSTRATION ONLY
Y N
Machine learning identifies the drivers of the final underwriting decision Which fields on your application form best predict your final underwriting decision? Which medical evidence could be eliminated without a risk cost?
Sample application form
Y N
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Var30 Var29 Var28 Var27 Var26 Var25 Var24 Var23 Var22 Var21 Var20 Var19 Var18 Var17 Var16 Var15 Var14 Var13 Var12 Var11 Var10
Var9 Var8 Var7 Var6 Var5 Var4 Var3 Var2 Var1
Relative importance
Varia
bles
Ranking the importance of case data, base questions and disclosures
FOR ILLUSTRATION ONLY
Impact of reducing underwriting variables on predictive power
25% 75%
*Area under the curve (AUC) is a measure of predictive power. A model that randomly decides the underwriting decision would produce an AUC of 0.5 and a good model over 0.85
AUC*
Median
% o
f all
ques
tions
incl
uded
100%
75%
60%
50%
25%
0%
0% 10% 20% 30% 40% 50% 60% 70% 80% 90%
100%
Score (probability of accepted standard in the final underwriting decision)
% o
f cas
es
STD Non-STD
How do risk implications arise from streamlining underwriting?
Based on testing set after building the model on the training data
Model score for the probability that a case is accepted standard in the final underwriting decision
• White area indicates cases predicted by the model as “accepted standard” but should be non-standard
• Also called false positives, i.e. cases “slipping through” incorrectly on standard rates
• Risk implications have to be managed
• The score is translated to an underwriting decision using a threshold above which an “accepted standard” decision is given
1 0
Score for a given case
Threshold
1 0
Risk management considerations Confusion matrix
Modelled underwriting decision
Standard Non-standard
Actual underwriter decision
Standard
Non-standard
A B
C D
A
B
C
D
True positive
False negative
False positive
True negative
Sub-standard loadings
Analysis of false positives • Conditionally accepted • Declined • Exclusions
Risk management controls • Knock out criteria • Non-disclosure • Monitoring:
• New business mix • Random holdout • Claims
• Ability to re-train models
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0% 2% 4% 6% 8% 10%
Risk implications of streamlining underwriting Trade off between STP, number of underwriting questions and the score threshold
Current STP
50% STP 100 variables 0.8% “slip-through”
80% STP 100 variables 2.6% “slip-through”
Number of underwriting variables
10 100 All (1000) S
TP
% of cases incorrectly predicted as standard
Risk implications of streamlining underwriting Using actuarial, underwriting and claims expertise to understand risks posed by false positives
Other conditions?
% o
f slip
thro
ugh
case
s
Declined Postponed Conditionally Accepted
No loading
Loading
Loading
No loading
No
excl
usio
n E
xclu
sion
Breakdown of false positives Conditionally accepted slip through cases Pricing assumptions
?
?
?
?
Risk implications of streamlining underwriting Trade off between STP, number of underwriting questions and the score threshold
STP
0%
10%
20%
30%
40%
50%
60%
70%
80%
90%
100%
0% 2% 4% 6% 8% 10%
Current STP
50% STP 100 variables 0.45% loading
80% STP 100 variables 1.5% loading
Percentage loading to qx (best estimate)
Number of underwriting variables
10 100 All
Deployment process and considerations Integration with Munich Re’s automated underwriting engine
Agent application
§ Integration of predictive model with underwriting engine and back office systems
§ Availability of model inputs at the required stages
§ Offline/online agent use of engine
§ Knock out by product/risk types, demographics, etc.
§ Monitoring of model performance and risk implications
§ Model maintenance
Online Offline
One decision
Explaining a single AI underwriting decision
Var 1 Var 2 Var 3 Var 4 Var 5 Var 6 Var 7 Var 8
Var 9 Var 10 Var 11 Var 12 Var 13 Var 14 Var 15 Var 16
Var 17 Var 18 Var 19 Var 20 Var 21 Var 22 Var 23 Var 24
...
... …
… … … … … …
… … … … … … … …
Q6
Total shap value
Shap value*
… … BMI
Reducing / replacing traditional underwriting with external data
3
Uses of bank and retail data
§ Frictionless underwriting: § Identify which customers are eligible for reduced underwriting and minimal / no price increase
§ Assess the degree to which bank data can replace some/all traditional underwriting which enables an improved customer experience, increased take up and reduced cost of underwriting
§ Increased STP (automation): predictive models could avoid many referrals
§ Propensity to buy: § Enables targeted digital offers with increased take up
§ Bank customer segmentation enables product offers that match customer needs
§ Competitive pricing: less (friction in) underwriting with no / minimal price increase
§ Data-driven product solutions: § Clustering of bank customers creates segments for matching / designing relevant products
§ Combined bank and insurance products that evolve with customers’ life stages (events) and are triggered and streamlined using bank and other data
Considerations in achieving frictionless underwriting Using bank data can potentially avoid a price increase when reducing traditional underwriting
Price
Level of underwriting friction e.g. Number of questions
GIO (Very little is known about the customer)
SIO (Little is known about the customer)
Fully underwritten (A lot is known about the customer)
Better than fully underwritten (A lot is known about the customer and incorporates bank data)
76% 78% 80% 82% 84% 86% 88% 90% 92%
Top 5 Top 15 Top 20 Top 40 Top 100
Acc
urac
y
Number of underwriting questions (As proxy for underwriting friction)
GIO
SIO Full u/w
Replacing traditional u/w with bank data (A lot is known about the customer)
Thank you Lee Sarkin Head: Regional Analytics Centre (Life & Health) Asia-Pacific, Middle East and Africa [email protected]