SPSS - UK Customers

Provident Financial

The UK’s leading home credit provider detects and prevents fraud with predictive analytics

The situation

Fraud detection and prevention is important to Provident Financial plc’s Personal Credit division, Provident Personal Credit (PPC). A leading lender in the United Kingdom’s home credit industry, PPC’s agents make and collect cash loan payments in person.

The critical Issue

In the personal loans business, cash changes hands daily in thousands of small, door-to-door transactions, which provides many opportunities for fraud – for example, stealing loan funds or manipulating collections to increase bonuses. Fraud is a harsh fact of life, but the number of opportunities available to commit fraud, combined with the difficulty of tracking cash exchanges, makes it difficult to detect. As well, strict UK consumer protection laws and rules of evidence also make it difficult to prosecute fraud – even when it is uncovered.

With more than 25,000 agents serving 1.5 million customers in the UK and Central Europe, PPC needed an effective system to detect and prevent fraud.

The solution

Using IBM SPSS Statistics Base* and IBM SPSS Modeler*, PPC’s investigators were able accurately to identify and prosecute fraud, leading to annual reductions in losses.

…Nearly 80 percent of the fraud we identify is found using IBM SPSS Statistics Base and IBM SPSS Modeler.
- Paul Wilkinson-Smith
Fraud Analyst

The results

The frequency and magnitude of agent and customer fraud was reduced
To help reduce fraud and target offenders, PPC used IBM SPSS Statistics Base and IBM SPSS Modeler. With these comprehensive data analysis and data mining programs, it narrowly focused its investigative efforts and identified agents who should receive ‘targeted visitations’ from fraud investigators.

According to Paul Wilkinson-Smith, a fraud analyst in PPC’s Field Security Department, most losses were the result of fraud by individual agents; however, in some cases agents, management and customers could conspire to steal funds. “With IBM SPSS Statistics Base, we were able to create a very effective statistical profile that tracked fraudulent activities back to our agents,” he said. PPC used a model built with IBM SPSS Statistics Base – using an analytical technique called logistic regression – to evaluate about 10 million customer records in its Sybase® database each week. The model runs on the same HP UNIX™-based server as the Sybase system, and generates visitation forms that are distributed to 60 area security managers for follow up. Security managers use these leads to conduct audits of agents and customers.

Money saved through early fraud detection
“The combination of a professional investigative field force and our IBM SPSS Statistics Base and IBM SPSS Modeler systems definitely has reduced the incidence and magnitude of our theft problem, as we detect and/or deter it earlier,” Wilkinson-Smith stressed. “In fact, nearly 80 percent of the fraud we identify is found using SPSS modeling technology.”

Wilkinson-Smith said PPC used a combination of its own business knowledge, IBM SPSS Modeler neural networks, and rule-induction models to further refine the data it extracts with the IBM SPSS Statistics Base logistic regression model. “When we first brought in IBM SPSS Modeler, we had already had many wins using IBM SPSS Statistics Base to profile fraud activity,” he explained. “We wanted to use IBM SPSS Modeler to help us profile fraudulent customers. Initially, consultants helped us develop models that would work for us, and today, using IBM SPSS Modeler and IBM SPSS Statistics Base together, we have the two systems complementing each other very nicely.”

“We split our test data in half,” Wilkinson-Smith explained. “We used half to train the IBM SPSS Modeler neural network and the other half to test the network once we believed it was ready. We also tested the network using raw data that had not been analysed previously. Then we looked at the neural network results and manually evaluated that data to determine how well the network had performed. We tweaked the neural network to some extent by adjusting thresholds, maximum performance levels, and other factors. While the neural net is self-contained, it can be adjusted and tuned for optimal performance.”

Investigators’ time was saved and the prosecution rate increased
The model further enhanced the company’s fraud-detection capability. “We can look at the rules that IBM SPSS Modeler generates and test them against known data,” Wilkinson-Smith said. “This combination of modeling tools provides a very powerful data mining system that enables us to target our investigations and save our field agents an unbelievable amount of time.”

“Fraud profiles change very frequently,” he added. “IBM SPSS Statistics Base and IBM SPSS Modeler enable us to change our models as often as we need to. We can retrain the neural network and rule-induction engine whenever we detect a shift in fraud patterns and profiles. I am a big IBM SPSS Modeler fan. It provides a superb visual development environment that saves us time and money. It’s brilliant!”

Interested in how you can reduce fraud and control losses? Download the Provident Financial PDF here.

IBM SPSS Statistics Base* and IBM SPSS Modeler*, formerly called SPSS® Statistics Base and Clementine®, are part of SPSS Inc.’s Predictive Analytics Software portfolio.