Resource Data

A data-driven approach to detecting cash benefit theft

A case study explaining how a predictive, data-driven machine-learning model was developed to detect unauthorized cash benefit withdrawals more quickly and accurately in California.

Organizations: California Office of Data and Innovation (ODI), California Department of Social Services

Authors: Monica G. Bobra, Ian Rose, Aeri Shan, Jason Lally, Kimberly Hicks, Esa Eslami, Konrad Franco, Joaquin Carbonell, Batool Hasanzadeh, Peter Amerkhanian, J. Steven Raquel, Neelam Joshi, Daniel Rodda, Brenda Moran

Published Year: 2026

View Source

Download Resource

This resource describes how analysts used transaction and reimbursement data to train a random-forest model that identifies suspicious Electronic Benefits Transfer (EBT) cash withdrawals, addressing challenges with data access, imbalanced classes, and feature engineering.

The model successfully flags theft with about 82 percent accuracy, reduces the time to identify theft from months to days, and enables staff to focus on targeted prevention and intervention efforts. By automating data pipelines and embedding geolocation and temporal features, the approach enhances capacity for real-time monitoring and supports future extensions to other types of fraudulent benefit activity.