aaim14 ahsley big data big opportunities · 2014. 10. 9. · microsoft powerpoint -...

10/7/2014

1

Proprietary and Confidential | © General Reinsurance Corporation

Big Data | Big Opportunities


AAAI – Association for the Advancement of Artificial Intelligence

AAIM | September 2014

“Predictive analytics is concerned with the prediction of future probabilities and trends based on observed events”

PA measures

quantitative effect

of multiple simultaneous

characteristics for a

defined outcome

2

10/7/2014

2


Applications

Inputs

– Time of day

– Day of week, year

– Weather

– Variability in speed, spacing

Traffic



Applications

The U.K.’s Royal Shakespeare Company used

analytics to look at its audience members’ names,

addresses, performances

attended and prices paid

for tickets over a period of

seven years. The theater

company then developed

a marketing program that

increased regular attendees

by more than 70% and

its membership by 40%.

Marketing


10/7/2014

3


Applications

Spam filter

Medicine

• Framingham Risk Index predicts cardiac events

• Cancer staging predicts mortality and recurrence

Amazon and Netflix suggestions



Mechanics


AssembleDescriptors of prior events insurance applicant –

age, sex, tobacco, blood pressure, cholesterol, socioeconomic status, consumer purchase records.

Identify Descriptors that correlate with outcome (mortality).

IntegrateEffective predictors and interactions between predictors in prediction algorithm (model).

RelateValue of algorithm result to mortality risk (A/E; underwriting class).

6

10/7/2014

4


Marketing

Product Design

Underwriting and New Business Process

Gen Re Life Insurance Market Survey

Potential

Uses of

Predictive

Modeling




Barriers To Predictive Modeling


10/7/2014

5



How Prevalent is Predictive Modeling



Products in U.S. Insurance Market


● Biomedical

– Conventional underwriting

evidence data(lab, paramed exam)

● Consumer behavior

– Unconventional data

10

10/7/2014

6


Biomedical

Industry lab vendors

– CRL SmartScore, ExamOne Risk IQ

– Mine historical customer results

– Use SSDMF to infer mortality outcome

BioSignia

– Digest clinical literature on relationship between underwriting evidence and mortality

– Derive relationships between each parameter and mortality risk

– Synthesize results across many studies into unified mortality risk equation



Consumer Behavior

Deloitte Consulting

Ignore conventional underwriting evidence

Mine electronic databases of consumer history

– Credit card purchases

– Warranty registration

– Survey responses

Relate this profile to risk of disease and mortality

Hundreds of parameters available for inclusion in model

Construct unique model for each client company

– Choice of parameters to include / exclude

– Tune to customers of each company


10/7/2014

7


Insurance Industry Labs Common Threads

Dataset of all lab customers who applied for insurance in past 10-15 yr

Even Heritage with smallest market share has millions of records

Datapoints all tests run plus all paramed measurements recorded (more recent and less prevalent than test results)

Social Security Death Master File to ascertain vital status and date of death

Statistical analysis to quantify overall mortality risk

Aggregate of all applicants may not fit any one company, but material biomedical differences unlikely



Lab Models Common Weaknesses

Applicants only, no knowledge of underwriting result, medical history

Model inaccurate to extent that lab data duplicates known medical risk (unless use model as substitute for other underwriting)

SSDMF incomplete, inversely with age


10/7/2014

8


Upside

Biomedical

– Multiple criteria for preferred considered separately distorts overall measure of risk

– PA multiple simultaneous variables can yield • More efficient risk classification• Less overlap among risk classes • Recognition of interactions that represent different risk than sum of the parts

Deloitte

– Faster, cheaper, automated underwriting without need for blood, urine, exam

Current uptake

– CRL Smart Score, ExamOne RiskIQ and Deloitte have active users



Validation

Industry labs

– Demonstrate that score corresponds to mortality experience – more later

Deloitte

– Demonstrate that score corresponds to risk class assignment from existing underwriting process

– Replication of underwriting action immediate – no need for experience to develop or retrospective study


Mo

rta

lity A

/E

SCORE

Conventional UW Class

De

loitte

in C

lass

1 2 3

1

2

3

16

10/7/2014

9


Examples


Fac underwriting assessment

>40,000 cases per year

Placement rate low teens

How can we make more successful offers on the cases we don’t write?


10/7/2014

10


Facultative Mortality Model

Compare mortality between placed / not placed

Is underwriting / pricing efficient?

– Degree of risk

– Impairment

– Other variables

Discover where we have room to make more offers / better offers

– Declines are less risky than we expected

– Convert underwriting assessment to multivariable function



SSDMF Accuracy

Mortality of unplaced cases is invisible

Use SSDMF to infer deaths

Comparison to in-force mortality experience

Measure accuracy of SSDMF against Gen Re claims


Byproduct of facultative unplaced analysis

In-force All deaths observed

Unplaced Incomplete reporting, but by how much?

20

10/7/2014

11


SSDMF Accuracy


Gender

75%74%21


SSDMF Accuracy


Age at Death

40%

45%

50%

55%

60%

65%

70%

75%

80%

85%

90%

0-19 20-29 30-39 40-49 50-59 60-69 70-79 80-89 90+

22

10/7/2014

12


Implications

Claim analysis allows us to adjust for undetected deaths in

Facultative unplaced analysis

Unclaimed property

Annuity surveillance

Industry lab vendor mortality score model construction



Lab Score Validation Project

Hypothesis

– Refine preferred / STD risk and reclassify more consistently

– Qualify more applicants or adjust prices for risk classes

Demonstrate efficacy (correlation between each product

assessment and observed mortality)

– Direct company could implement it

– Reinsurer could reflect it in pricing

– Regulator / producer could accept it

Lab vendors derived model from insurance applicants / SSDMF

How does it perform on underwritten population?


10/7/2014

13


Predictive analytics for medical directors

Enormous potential for risk classification

Make your company an informed consumer of

vendor products

Generate ideas and guide construction and

implementation



Thank You

aaim14 ahsley big data big opportunities · 2014. 10. 9. · microsoft powerpoint -...

Documents