Transcript Innovations in Detecting Suspicious Claims

Innovations in Detecting Suspicious Claims
MEASURE, MANAGE, & REDUCE RISK
1
SM
Agenda
• Impact of insurance fraud
• Resisting fraud effectively
• Building fraud detection solutions
– Keep up with changing scams
– Maximize value from structured data
• Business rules
• Predictive modeling
– Leverage textual data assets
– Exploit claim networks
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
Why Focus on Fraud?
• It is a big problem
–
of personal injury claims contain elements of
26%
fraud1
– $50 to $100 of policyholder premiums go to pay
fraudulent claims2
• It is widespread
– Fraudsters operate across touch points and verticals
– New entrants driven by the economy
• It keeps changing and morphing!
1 2001
study conducted by the Insurance Bureau of Canada
2 http://www.infoassurance.ca/en/preventing/automobile/fraud.aspx
Resisting Fraud Effectively
• Corporate culture
– Fighting fraud must be a core responsibility
– Organizational measurements must be aligned
• e.g., fraud investigation impact on cycle time
• Effective process
– Effective antifraud training programs
– Well-defined processes for detection, referral, and investigation
– Integration with technology/solutions
• Systematic fraud detection solutions
– Best-in-class solutions that evolve to stay current
– Multiple techniques to cover different angles and types of data
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
Building Fraud Detection Solutions
1
Understand
Fraud red flags,
schemes, and
scams
5
Evaluate
Build
SIU investigation
and feedback on
evolving scams
Systematic fraud
detection
mechanisms
4
Refer
Score
Business thresholds
to refer claims to
SIU
Process to score
claims for fraud
potential
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
3
2
Example Scams
• Staged auto accidents
– Swoop-and-squat – Car in front of you stops suddenly
– Wave-on – claimant indicates it is safe for you to merge or pull out of a
parking space, but then runs into you
• Repair shop scams
– Airbag fraud – bill for new airbags but replace with stolen or salvaged
– Burying the deductible – inflate estimates to make insurer pay the
deductible (collusion with insured)
• Owner give-ups
– Owners report their used car stolen and then set it on fire.
Total loss ensures insurance pays off the entire car loan
• Auto glass fraud
– Bill for a windshield replacement when only a chip repair was done
– Soliciting glass claims
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
Scams Change and Evolve
• Increasing PIP fraud
• Rise in property
scams (e.g., hail)
• Effects of the new
economy
– Auto give-ups
– Glass claims
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
Fraud costs in Ontario top those in other parts
of the country… according to panelists at an
RBC Insurance roundtable on fraud.
Those costs represent an estimated $1.3
billion of $9 billion in premiums in the
province, the insurance executives noted
during the July 28 [2010] discussion…
The average cost of a claim in Ontario rose
from $30,000 in 2005 to $53,000 in 2009,
according to Insurance Bureau of Canada (IBC)
data. That’s markedly more than average
claims costs in Alberta ($3,689) or Nova Scotia
($5,904).
Changing Scams
Source - NICB ForeCAST Report - 3Q Referral Reason Analysis (Ann Florian, Strategic Analyst )
MEASURE, MANAGE, & REDUCE
USING STRUCTURED DATA
Structured Data in Claim Systems
• Policy details
– Insured details (age, sex, etc.), # of years insured,
policy inception date, etc.
• Loss details
– Date and time of loss, location of loss, details of vehicles
involved in loss, etc.
• Claimant details
– # of claimants, injuries, treatment dates and amounts
• Representation
– Attorneys involved (if any), date of engagement, etc.
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
Business Rules: SIU Scorecard
Red Flag / Indicator
Points
Insured reports accident did not happen
100
Informant notifies carrier of suspected fraud
100
Unexplained inconsistent damages
100
Indication that the accident was a setup
100
Claim reported more than 20 days after loss
40
Minor impact
30
Loss within 90 days of a new policy
20
Multiple injured claimants
30
Unrelated claimants with same doctor
25
Unrelated claimant with same attorney
25
Treatment started over 15 days after injury
30
Claimant had another BI claim
40
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
Scoring & Referral
1. For each claim,
determine
indicators that
apply
2. Add the
corresponding
points
3. If total points > 99,
refer to SIU
Predictive/Statistical Modeling
• Supervised models
– If target flag (suspicious/not-suspicious) tags are available
on a historical body of claims
– Many model forms available
• Naïve Bayes models
• Decision trees
• Logistic regression
• Neural network classifiers
• Etc.
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
Decision Tree for Fraud Detection
Clmt Vehicle
= OlderAmerican
(70%)
Insd Vehicle
= Luxury
All Claims
(Fraud Rate
2%)
Insd Driver =
Female
(25%)
# Clmts > 1
(10%)
(5%)
# Clmts = 1
Insd Driver =
Male
Insd Vehicle
= NonLuxury
(1%)
(3%)
= Refer to SIU
MEASURE, MANAGE, & REDUCE
= Alert adjuster
Clmt Vehicle
= OlderJapanese
(45%)
Clmt Vehicle
= Newer
(7%)
= Settle claim
(10%)
TEXT MINING
FOR ADDITIONAL LIFT
Text Mining Adjuster Notes
IT APPEARS THAT THIS WAS A LOW-IMPACT COLLISION WHERE THE INSURED’S
FOOT SLIPED OFF THE BRAKE, AND SHE ROLLED INTO THE REAR OF THE CLAIMANT.
THIS IS CONSSTENT WITH THE FACT THAT THERE WAS NO PROPERYT DAMAGE
CLAIM MADE TO THE CLAIMANT VEHICLE. UNDER THE CIRCUMSTANCES, HOW THE
CLAIMANT COULD HAVE SUSTAINED SUCH SEVERE SHOULDER INJURIES AS A
RESTRAINED DRIVER APPEARS RATHER SUSPECT.
Questionable
Injuries
Low Impact
Exaggerated
Treatment
NO PROP DMG FOR INS AND CLMT AS COLL HIT WAS LOW. CLMT CLAIMS INJ
FROM AX AND TRTD W CP AND PT EXTENSIVELY. TX APPEARS EXAGGERATED.
MEASURE, MANAGE, & REDUCE
Unique Insights in Text
INSD R/E CLMT VEH WHEN IT BRAKED SUDDENLY NEAR HIGHWAY EXIT. INSD
THINKS SPEED OF TRAVEL ABOUT 25 MPH. INSD SUFFERED AIRBAG BURNS.
MULTIPLE CLMTS IN VEHICLE WERE INJ BUT WAIVED AMBULANCE.
Insured R/E
Claimant
Near Highway
Exit
No EMR and/or
Ambulance Waived
• “Structurized” data
– Structured fields created with codes/values extracted using
text mining, e.g.:
• Near Highway Exit = Y/N
• Low Impact = Y/N
MEASURE, MANAGE, & REDUCE
Better Detection with Text Mining
Clmt Vehicle =
Older-American
(70%)
Insd Driver =
Female
# Clmts > 1
(5%)
(10%)
Insd Vehicle =
Luxury
Clmt Vehicle =
Older-Japanese
(25%)
(45%)
All Claims
Highway
Exit = Y
(Fraud Rate
2%)
(3%)
(15%)
(1%)
= Refer to SIU
MEASURE, MANAGE, & REDUCE
(10%)
(7%)
Insd Driver
= Male
# Clmts = 1
Clmt Vehicle =
Newer
Insd Vehicle = NonLuxury
No EMR = Y
(50%)
Low Impact
=Y
Exaggerated
Treatment = Y
(5%)
(40%)
= Alert adjuster
= Settle claim
MINING NETWORK DATA
Industry Data: ISO ClaimSearch®
Casualty
•
•
•
•
•
•
•
•
•
•
•
•
Workers Compensation
Automobile Liability
Medical Payments
Personal Injury Protection
Auto Medical Payments
Homeowner’s Liability
General Liability
Disability
Personal Injury
Employment Practices
D&O / E&O
Fidelity and Surety
>170 Million
•
•
•
•
•
•
•
•
•
•
Property
Homeowners
Farm Owners
Fire
Allied Lines
Commercial
Ocean Marine
Inland Marine
Burglary and Theft
Credit
Livestock
>36 Million
Auto
• Theft Claims
• Theft Conversions
• Vehicle Claim System
(damage estimates from
vendors)
• Shipping & Assembly
• Salvage Records
• Impound Records
• Export Data
• International Salvage and
Thefts
>395 Million
Insurers representing 93% of direct written premium, National Insurance
Crime Bureau, and law enforcement agencies
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
Querying Claim Networks
ISO’s NetMap tool for link analysis and visualization
MEASURE, MANAGE, & REDUCE
Characterizing Network Measures
Density
Centrality
MEASURE, MANAGE, & REDUCE
Betweenness
ORA (Organizational Risk Analyzer) from the Center for the Computational Analysis of Social and Organization Systems at CMU
Network Measures Add Value
Clmt Vehicle
= OlderAmerican
= Structured data
(70%)
= Text-mined data
= Network data
# Clmts >
1
All
Claims
Insd
Driver =
Female
(10%)
(Fraud
Rate 2%)
# Clmts = 1
(1%)
Clmt Vehicle
= OlderJapanese
(45%)
(80%)
(25%)
Clmt
Vehicle =
Newer
Density
= Med
(10%)
(40%)
Insd Vehicle =
Non-Luxury
(7%)
(5%)
Insd Driver =
Male
Highway Exit
=Y
(3%)
(15%)
Low Impact =
Y
Exaggerated
Treatment = Y
(5%)
(40%)
M E A S U R E , M A N A G E , & R E D U C E R I S K SM = Refer to SIU
Density
= High
Insd
Vehicle =
Luxury
= Alert adjuster
No EMR = Y
(50%)
Density =
Low
(2%)
= Settle claim
Summary
• Undetected fraud impacts the bottom line
• Effective fraud detection requires
– Corporate focus
– Process and training
– Effective tools and solutions
• Good solutions exist, but there is more to come
–
–
–
–
Cross-vertical fraud detection
New data sources (LPR data, cell phone data, etc.)
Geospatial data and technology
More innovations with predictive modeling, text mining, and
network mining
M E A S U R E , M A N A G E , & R E D U C E R I S K SM
Feedback and Questions
• Send feedback to:
– Janine Johnson
– +1.415.276.4105
– e-mail: [email protected]
M E A S U R E , M A N A G E , & R E D U C E R I S K SM