Predictive Modeling and Behavioral Analytics in Unemployment Insurance
Meet Carol. She recently got UI Overpayments Cost $3.2B Annually laid off from her job and has Data from Internal and External Sources decided to apply for UI benefits Objective: Identify and prevent improper payments in Data Sources: unemployment insurance benefits administered • Historical claims data (2012 – 2013) — Client’s Unemployment Framework for Claims Background: According to the USDOL, improper payments cost and Tax Services (uFACTS) database more than $3.2B to Unemployment Insurance (UI) trust funds in 2013, over 8% of the $40B in benefits paid out. Currently, This is a pointer style • ZIP code level demographic data (2013) — overpayments are often identified after payments have already been Easy Analytic Software, Inc. (EASI) made – potentially months later – at which point it can be difficult to Number of claims at the address, recover the costs. Data Types Whereand Sample Variables Claimants Live geographic area, Past fraud in the region UI Overpayments by Cause What Claimants Report Gender, Age, Occupation
To initiate her benefits, Carol Timing of filing claims, Method of
19% Benefit Year Earnings How Claimants File Claims filing, Changes in established fills out the 34% Separation Issues Initial Application behavior Number of base period Work Search Where Claimants Work employers, Past fraud at the 33% employer, Employer attributes 14% Other Past fraud and overpayment As she fills out the application, Past History on Claimants behavior Carol’s risk of getting overpaid is assessed and her application is customized Models Predict Fraud in Real-Time Benefits Realized through Model Applications Three logistic regression models were developed for different Prioritization of Investigatory Efforts overpayment types. Predictive strength, data availability/reliability, Carol completes and submits to maximize staff efficiencies and identify improper payments and business interpretation were all considered during variable the application faster. Workload on adjudication staff is reduced by 30-40% with selection. Higher Lower minimal impact on overpayments by: Applicant Information Font Color: Risk Font Size: Importance State agency completes fact- Applicant Name: Carol Jones finding and adjudication • Prioritizing high risk claims for Weeks Certified: 10 Benefits Paid: $5,000 investigation Risk Assessment: 95
To receive benefits, each
• Auto-adjudicating low risk Priority level: HIGH week Carol fills out the claims Weekly Certification Survival Curve: Audited vs. Non-audited Claimants Enhanced Work Search Audits to help claimants get back to work up to 10 weeks faster. Based on behavioral changes Non-audited claimants stay on Separation Overpayments Model Benefit Year Earnings Fraud Model and new information reported, (Initial Application) (Weekly Certification) Carol’s risk of committing benefits longer – decrease fraud is assessed and her average time on benefits by Model Performance application is customized conducting automated work 5% Highest Predicted Overpayment search audits, especially for high 300% • 2x higher rate of fraud at initial claim • >3x higher rate of separation overpayment risk claimants. • Weekly certs: >4x higher rate of fraud 250%
Carol completes and submits
Behavioral Messaging Actual Relative Fraud / Overpayment
200% the certification
to prevent improper payments by promoting honest behavior. 150% 10% Lowest Predicted Overpayment Serving dynamic and customized messages to 10% of the highest • 89% lower rate of fraud at initial claim • 71% lower rate of separation overpayment risk claims responsible for 49% of earnings fraud can change 100% • Weekly certs: 89% lower rate of fraud claimants’ behavior by the time they hit “Submit”: State Agency completes fact- Reminder 50% finding and adjudication • Certification Boxes — 9 out of 10 people in your county report earnings accurately. If you 0% “I certify that the information worked between 2/24/2013 and 3/2/2013, please ensure you report these earnings. Please report all information accurately. For each certification week I am providing is accurate.” Edit Responses Continue
-50% that Carol meets the eligibility • Persuasive Messages —
Earnings Fraud (Initial Application) requirements, she is paid Separation Overpayment (Initial Application) “We verify your employment and earnings information.” -100% Earnings Fraud (Weekly Certification) Benefits Top • Educational or clarifying messages — 1 2 3 4 5 6 7 8 9 10 5% “To qualify for benefits, you must have been physically able to Low Highest In a few weeks, Carol finds a Model Risk Assessment from the work, meaning you were not injured, ill, or otherwise physically Baseline new job and is able to roll off Overall Average Fraud / Moment Claim/Certification Submitted UI benefits! unable to accept work.” Overpayment Rate
Anomaly Detection Identifies ~50% of Hard Results Measured with RCTs
Fraud at Time of Application After implementation of the solution, a series of randomized control trials (RCTs) will be run to quantify the following expected effects: An ensemble of anomaly detection algorithms was developed to combat organized fraud schemes like identity theft and fictitious This is a pointer style
employers. • Improved ease-of-use and user experience
• Back to work, faster • Sentinel Effect – increased honesty under evaluation
• Reduced overpayment dollars
Cluster Analysis Association Rules • Decreased time to detection and resolution of overpayments
STATE • Efficient agency resource allocation
Predicted Age
• Earlier identification of emerging fraud schemes
As results develop over 10-12 weeks, continous monitoring of the
Benford’s Law Reported Age Likelihood Methods results is necessary in order to optimize risk-based business rules.