Document 7663424

Download Report

Transcript Document 7663424

Varieties Statistical Fraud Models:
30 Models in 30 Minutes
Daniel Finnegan, CFE
ISO Innovative Analytics
Quality Planning Corporation
Benford’s Law in Accounting Fraud
Odds of Obtaining as 1st Digit (%)
35
30
25
20
Odds of Obtaining as 1st Digit
(%)
15
10
5
0
1
2
3
4
5
6
7
8
9
Tests for Manufacture Numbers







Frequency or equidistribution test (possible elements
should occur with equal frequency);
Serial test (pairs of elements should be equally likely to be
in descending and ascending order);
Gap test (runs of elements all greater or less than some
fixed value should have lengths that follow a binomial
distribution);
Coupon collector's test (runs before complete sets of
values are found should have lengths that follow a definite
distribution);
Permutation test (in blocks of elements possible orderings
of values should occur equally often);
Runs up test (runs of monotonically increasing elements
should have lengths that follow a definite distribution);
Maximum-of-t test (maximum values in blocks of elements
should follow a power-law distribution).
IRS Audit Selection System
1964 Rule-Based Scoring System
1970’s TCMP Statistical Audit System
2003 NRP System:
A. Random Audits of Sample of Returns
B. Identification of Returns “In Need of
Examine”
C. Statistical Model of DIF score of “Probability
of Need to Examine”
D. Monitoring and Update of System
Text Mining for Fraudulent Medical Bills
Search for identical typos
 Search for identical prognosis
 Search for date discrepancies

 Holidays
 Claimant out of town/dead
Medical Usage Pattern Fraud Analysis





Uniformly high numbers of
treatments (Normed on Diagnosis)
High number of modalities per
treatment
Few Patients Recover Quickly
Low Percentage of Objective Injuries
Treatment Ends Abruptly at Payment
of Claim
FAIS Money Laundering Statistical
Detection




Link Analysis with Known Criminal
Elements
Pattern Analysis such as Large Sum
Deposited and Immediately
Withdrawn
Benford Distribution of Deposits and
Withdrawals
Circular Movements of Funds
Network Analysis of Auto Accidents
Daniel
Glenn
Richard
Staged Accident Ring
Sequential Handling of Questionable
Claims


Random Sample of 3,000 BI Claims
Decision Flow Model
Adjust and
Settle
Low
Initial
Review
Fraud
Score 1
Middle
High
Clear
Questions
Fraud
Score 2
SUI
Timing Claims Curves
Claims by Policy Week
30
25
Claims
20
15
10
5
0
1
3
5
7
9
11
13
15
17
19
21
23
25
27
Week
29
31
33
35
37
39
41
43
45
47
49
51
Other Threshold Fraud Models
Adding Coverage for Comp
 Two-Year New Vehicle Replacement
 School Lunch Eligibility

Deviant Purchase Patterns for Credit
Card Fraud




Identification of Individual Purchase
Patterns (Neural Net Models)
Identification of Typical Fraud Purchase
Patterns (Electronics, International
Spending)
Movement out of Typical Toward Fraud
Patterns
Expert Patterns Such Geographic
Dispersion of Purchases
Geographic Analysis of Staged Accidents
Insured
Claimant
Accident
Attorney
Chiropractor
Chorpo
Geographic Analysis of Staged Accidents
Insured
Claimant
Accident
Attorney
Chiropractor
Chorpo
Driver’s License Translator Fraud

Pass Rate:
 51% vs 95+%
Time to Complete
 30-60 Minutes vs 10-15 Minutes
Accidents by Time Since License
200
180
160
140
Accidents

120
Translator
Matched
100
80
60
40
20
0
1-6
7-12
13-18
19-24
25-30
Months
31-36
37-42
43-48
Insider Stock Dealing



MonITARS: Fuzzy Logic, Neural Nets,
Genetic Algorithms for London Stock
Exchange
Advanced Detection System (ADS) for
Nasdaq matches rule-based sequential
trading patterns
SONAR matches wire stories to stock
trading using pattern analysis to detect
stock manipulation
WC Premium Audit Selection Model






Statistical Modeling of 4 Years of Audit
Results
Holdback of 5th Year of Results
Combined Expert Theory and Inductive
Modeling
Final Model Built with Multiple Statistical
Methods:
 Decision Trees, MARS, GLM
Model Concentrated on Key Ratios by
Industry
Results more than Doubled Audit Returns
University Student Aid Fraud



Very High and Similar Hardship
Deductions (High Medical Bills)
Identical Applications for Student
Financial Aid (High Aid with No
Audit)
Fraud Clusters by Successful Sports
Teams
Work Load Analysis of Medical Billing
Fraud




Psychiatrist billing 80 hour work
days
Billing on 365 day years
Billing from distant locations
Billing for 200 patients per day
Adjuster – Vendor Pairing Models


Billing Pattern Analysis for 5 Million
Claims and 12 Million Payments
Dozen Questionable Patterns
Identified:
 Relative High Payment Average for
Adjuster and Vendor
 Identification of Vendors with Multiple
Payments to PO Box with Single
Adjuster
Social Security Disability Model



Random Sample File Review
Identified Decision Errors/Fraud
Built Multiple Models





Econometric
Decision Trees, GLM, Hybrid
Rule Violation
Decision Maker Focused
Final Artificial Intelligence Model
Sales Agent Rating Models

Sales Agents Mileage Model
 Low to Expectations
 Below Rating Cut Points

Missing Drivers
 Teenagers Low to Expectations
 High Permissive Use Claims

Frequent Claims After Comp Added
Food Stamp Store Investigation System



Prior System Viewed as a
Success
Random Investigation of 2,000
Stores
Statistical Analysis of Discovered
Violations
Food Stamp Investigation Outcomes
Discovered Violations
90
80
70
60
50
Random Rate
40
Trageted
Investigations
30
20
10
35
33
31
29
27
25
23
21
19
17
15
13
11
9
7
5
3
1
0
VIPER System
Statistical Pattern Targeting
 Random Component for Updating
 Geographic Clustering Component
 Tripled Discovered Violations
 Doubled Investigator Productivity

Thresholding Cell Phone Accounts



6-8 Percent Cell Phone Costs Fraudulent
High Volume of Calls and Turnover of
Fraud Requires Rapid Response
Account “Thresholding” Process Used
 30-Day, Fraud Free, Norming Process
 Account Specific Expert Rules on Duration,
Location, Timing
 Calls Scored Statistical Distance from Norms
 Percent of Potential Fraud Calls Monitored
 Norms Constantly Updated
Identity Theft Scoring
Scoring System Includes Variety of Data
Matching and Pattern Analysis Variables

High Numbers of Credit Card or Cell
Phone Applications from Address

Identity Variable Conflicts

Mail Drop Address

Impossible SSN
Dead, Issued Before Born, Un-issued, Impossible
Statistical Adjuster Assignment Models
Review of Areas of Fraud Loss
 Identification of Best Practices for
Handling Questionable Claims
 Sample Investigation of Matched
Samples of 1,500 Standard Handling
and 1,500 Enhanced Handling
 Statistical Modeling of Handling
Gains

Statistical Adjuster Assignment Models
$1,800
Average per Exposure Cost by Claims History and Handling
Method
$1,612
$1,600
$1,356
$1,400
$1,261
$1,200
$1,000
Standard Handling
$800
Enhanced Handling
$638
$600
$400
$200
$0
Questionable History
Unexpectional
Common Elements of Successful
Statistical Fraud Control




Statistical Methods Selected to Fit
the Problem (One Size Does Not Fit
All)
High Input from Substance Area
Experts
Feedback Loop Evaluates and
Updates System
Strong Integration with Operations