According to the Nilson Report, payment card issuers, merchants, and their acquiring banks lose over $11 billion to fraud each year. These enormous losses are indicative of the challenges that risk departments face when it comes to creating and updating their fraud detection models. The ability to quickly identify clear correlations between transaction characteristics and patterns of fraud to keep pace with evolving fraud tactics is critical.
A major payments technology firm explored the use of Ayasdi’s machine intelligence software to evaluate its existing fraud detection models. It used the software to ingest over a million credit card transactions and analyze hundreds of variables that characterized these transactions.
The Ayasdi team was able to produce actionable results within six weeks. The software quickly visualized the regions with fraudulent transactions and produced a listing of the statistically ranked features that described these regions. The analysis also helped the firm visually compare the ground truth to the existing model’s predictions and zero in on systematic model performance issues (both false positives and false negatives). It also helped the firm identify the statistically significant features that characterize these regions of false positives and false negatives to improve their existing models and create new ones. This approach increases fraud detection accuracy, reduces revenue losses, and improves customer satisfaction.