Harnessing Analytics to Minimize Debit-Card Fraud

Since it began providing Civil War soldiers with a secure means of sending funds back home, the U.S. Postal Service (USPS) has been one of the nation’s largest providers of money orders. Now post offices from coast to coast sell more than 85 million money orders, having a face value of more than $20 billion, each year. They accept payment for these money orders via cash or debit card only.

Today USPS point-of-sale systems are EMV-enabled, and all debit-card payments for money orders must use chip technology. Three years ago, however, these capabilities were not live. In October 2015, credit and debit card issuers in the U.S. shifted liability for fraudulent transactions to any merchant that failed to meet EMV specifications. It was an effort to push retailers and other vendors to adopt the new, more secure technology. Unfortunately, at the time, the Postal Service was not yet EMV-ready.

“We can’t disclose the exact size of the risk we were facing, but any loss is too great a loss for us,” says Elizabeth M. Schafer, treasurer of the U.S. Postal Service. “We’re an entirely self-funding entity, paying for our operations out of the sale of products like money orders. When debit card issuers shifted liability to us, we couldn’t afford to start paying for a lot of fraud losses.”

The treasury team determined that their best bet for dealing with the liability shift would be to repurpose a rules-based software engine that the Postal Service was already using to minimize fraud in its online sales. Schafer launched an initiative with that purpose, putting together a project team that included staff from corporate treasury; the enterprise analytics group’s advanced analytics team; the USPS’s Eagan, Minnesota accounting service center; and the Postal Service retail division.

The Eagan accounting service center pulled together a data set that included large numbers of recent money-order transactions—both fraudulent and legitimate—that the USPS had processed in similar locations and time periods. “We wanted our sample data to resemble something we would find in a normal transaction,” explains John Greaves, chief data scientist in the USPS’s advanced analytics division.

Next, they divided that data set into two. They set aside one subset of the data and ran the other subset through a machine learning tool designed to select the algorithm that most effectively predicted fraud. “The tool looked for patterns to discern the difference between a valid transaction and a fraudulent transaction,” Greaves says. “Our objective was to model the fraudulent cases that were shared with us. When we had the right algorithm, the software enabled us to develop discrete business rules that we could implement in a fraud-detection decision tree.”

Once the advanced analytics group had developed this model, they ran the other data subset through it to see how well the business rules held up against these new real-world examples. The model had “learned” well from its first data set and proved quite effective at predicting fraud within the second data set, even though it had never previously seen those transactions. In fact, two-thirds of all disputed charges in the validation-testing data set fell within the model’s top 2 percent of the riskiest transactions.

The Postal Service soon rolled out this rules engine in its day-to-day money order process. If a customer used a PIN debit card to purchase a money order in a post office, and if the transaction met certain criteria, then the transaction would be sent through the new rules-based engine to predict its fraud risk.

“The predictive model would use the business rules to determine whether to forward the transaction on to the card issuer for authorization,” Schafer says. “Transactions deemed too likely to be fraudulent wouldn’t be sent on for authorization.”

See also:

• Custom-built Model Drives Better Reserving for Liability Claims
• 2018 Alexander Hamilton Awards in Treasury Transformation

As a result of this system, the USPS was able to keep fraud losses in check, despite the liability shift. The system also achieved another goal, Schafer says: It did not hinder the Postal Service’s ability to serve customers who legitimately needed money orders. “It was important to make sure people who needed to purchase money orders to pay their bills weren’t harmed at all,” she says. “We’re confident we didn’t harm any of our ‘good’ customers because our help desk didn’t see an increase in contacts or a spike in the number of declined transactions.”

The treasury and analytics teams came at this initiative from very different perspectives, but their open communication facilitated the project’s success. “The data science and exploration group had never interfaced with finance before,” Greaves says. “Tapping their domain knowledge was extremely valuable. We took a quasi-‘agile’ approach to project management, completing tasks in two-week sprints and then checking back in with the project team for feedback on our progress. This enabled us to tweak very quickly if it turned out we were slightly off-track.”

Schafer concurs. “One secret to our success was our willingness to listen to one another,” she says. “Our groups had lengthy discussions about our different perspectives on what the data was telling us. We always respected one another’s expertise and opinion, which built a culture of collaboration and helpfulness that enabled this project to thrive.”

Within a year, the Postal Service had implemented EMV technology for processing debit cards, so it no longer retains liability for losses and no longer depends on its own predictive analytics to root out fraud. Nevertheless, the project has had lasting ramifications for both the treasury and the analytics teams.

“When the project was in progress, we regularly reported out to the CIO and CFO,” Schafer says. “Their involvement gave the project visibility throughout the Postal Service. And the more people have learned about our project—and heard about the great work the advanced analytics team has done here—the more they have started to reach out to the advanced analytics group to see how predictive analytics can help other areas of the organization.”

2019 Alexander Hamilton Awards

Has your organization completed an innovative initiative in treasury, finance, or risk management? We’d love to recognize your project with a 2019 Alexander Hamilton Award! Enter today!