Sequential fraud detection for prepaid cards using hidden Markov model divergence

- Novel stored-value card fraud detection system.
- Novel use of hidden Markov model for modeling merchant transactions.
- Findings on data from a large stored-value card transaction processor.
- Findings on experiments evaluating the model quality under various conditions.

Stored-value cards, or prepaid cards, are increasingly popular. Like credit cards, their use is vulnerable to fraud, costing merchants and card processors millions of dollars. Prior techniques to automate fraud detection rely on a priori rules or specialized learned models associated with the customer. Mostly, these techniques do not consider fraud sequences or changing behavior, which can lead to false alarms. This study demonstrates how a transaction model can be dynamically created and updated, and fraud can be automatically detected for prepaid cards. A card processing company creates models of the store terminals rather than the customers, in part, because of the anonymous nature of prepaid cards. The technique automatically creates, updates, and compares hidden Markov models (HMM) of merchant terminals. We present fraud detection and experiments on real transactional data, showing the efficiency and effectiveness of the approach. In the fraud test cases, derived from known fraud cases, the technique has a good F-score. The technique can detect fraud in real-time for merchants, as card transactions are processed by a modern transaction processing system.