Transcript Slide 1
Loss Prevention,
Auditing & Safety Conference 2009
Title Sponsor:
Improving Organizational Safety
Through Predictive Modeling
Kris Russell – Sr. Manager Risk Strategy
Insights, Research & Analysis
Wal-Mart Stores Inc.
Agenda
Introduction
Predictive Modeling Defined
Predictive Modeling in the Retail World
What to Expect
Summary/Conclusion
Q&A
Agenda
Introduction
Predictive Modeling Defined
Predictive Modeling in the Retail World
What to Expect
Summary/Conclusion
Q&A
Predictive Modeling Defined
Predictive Modeling – *Deloitte’s Definition
Data Mining
Algorithms
Segmentation
Segmentation
Vulnerable Store Identification
Focused Resource Deployment
*Deloitte Touche Tohmatsu
How Predictive Modeling Works
Multiple
Claim
Variables
Score
Predictive
Equation/
Calculation
Indicator
Score
How Predictive Modeling Works
Score Based Groups
Skill Matches Group
Better Initial Assignment
Predictive Modeling Concept – Example
Claim begins to exhibit traits that make it
suspicious to fraud investigators
Claim Investigation
Investigation Benefit
Traditional Fraud
Identificaiton Process
Fraud investigation
initiated
Claim is opened
When the model scores the
claim, it is flagged for
investigation
Investigation Benefit
Claim Investigation
Predicive Model Fraud
ID Process
Claim is opened
Early ID
Fraud investigation
initiated
Benefit
Prevention vs. Prosecution
Claim is closed
Claim is closed
Agenda
Introduction
Predictive Modeling Defined
Predictive Modeling in the Retail World
What to Expect
Summary/Conclusion
Q&A
Wal-Mart’s Predictive Modeling Philosophy
Combine Multiple Models
Produce Consolidated Score
Overall Claims Evaluation
Wal-Mart Litigation Model Case Study
Traditional Process
Random
Time Consuming
Goal: Flag High Potential Claims
Claim Opening + 30 Days
Identification
Claim Management
Wal-Mart Litigation Model Case Study
Uses 26 variables
The Question:
Science = Experience?
Outcome:
‘Lift’ in Identification
Wal-Mart Litigation Model Case Study
‘Lift’?
Traditional Approach ->
Total Claim Pool
(1,000's of
Claims)
Desired
Claim
Pool
Search Includes The Total Claims Data
Population
Total
Claims ID’d
Early
(% of Total
Population)
Predictive Modeling ->
Total Claim Pool
(1,000's Claims)
Desired
Claim
Pool
Narrow Sample
Search – Early
ID
More Cases
ID’d Early
Wal-Mart Litigation Model Case Study
Adjuster 1: 14 Years Experience
7 of 25
Adjuster 2: 25 Years Experience
6 of 25
Model
7.5* of 25
Agenda
Introduction
Predictive Modeling Defined
Predictive Modeling in the Retail World
What to Expect
Summary/Conclusion
Q&A
Predictive Modeling Life Cycle
Business
Understanding
Deployment
Evaluation
Data
Understanding
Data
Preparation
Modeling
Data Approach Choices
Decentralized vs. Centralized
Ownership of data
What to Expect
Data is Key
Ask For Help
Use an experienced actuary
Agenda
Introduction
Predictive Modeling Defined
Predictive Modeling in the Retail World
What to Expect
Summary/Conclusion
Q&A
Summary/ Conclusion
Predictive Modeling
Proactive Data Use
Improved Resource Allocation
Narrow the Window
Data is Power
Agenda
Introduction
Predictive Modeling Defined
Predictive Modeling in the Retail World
What to Expect
Summary/Conclusion
Q&A