ONLINE FRAUD DETECTION SYSTEM
Bank B is a major Australian bank with global presence. As most of banks, B processes about 2 million online transactions per day. Approximately 10~20 frauds might hide in the huge number of genuine transactions. An efficient fraud detection system is highly expected to protect their customers from fraudulent attack earlier, faster and more accurately.
B has been employing a rule-based expert system in online banking fraud monitoring for 10 years. Some prominent disadvantages are found in the existing rule system. First of all, high false positive rate was found especially where rules are manually tuned to cover newly emerging frauds. A rule is usually built to simulate risky scenarios, with which alerts are triggered on suspicious transactions. Some critical disadvantages are obviously found: (1) Balancing the detection rate and false positive rate are increasingly challenging due to the rapid development of fraud techniques by fraudsters. (2) Maintaining complex detection rules are costly and inefficient. The rules generated by the domain experts are considerably long so as to obtain good trade-off, as a result, they are hard to understand and maintain. (3) The rule tuning is usually difficult due to the emerging new techniques applied for fraudulent activities. (4) The expert system is a fake real time monitoring approach. In every 30 minutes, the transactions go through the checking process executed by detection rules in batch. A real time fraud detection system is essential to detect suspicious cases and recover the money instantly. (5) Refreshing detection rules has to be conducted manually via tuning the rules by domain experts, which is time consuming and low efficient.
To solve the problems challenging the real time fraud detection, we applied advanced behavior analytics and predictive modeling to build a real time detection system. Contrast patterns, which are rules matching fraudulent transactions with high matching rate and extreme small matching rate among genuine ones, were mined to filter genuine transactions thereby enormously cutting down the false positive rate. Automatic training process incorporates fresh transactions to upgrade the models with new knowledge guaranteeing a stable performance in detection. The monitor system upgrades the detection power by extracting emerging suspicious transactions into the retraining. The models can be easily refreshed on a daily/weekly basis. On the other hand, the algorithms are specially designed to automatically maximize the detection rate with specified low false positive rate. Upgrading the detection system can be conducted within a couple of minutes through an overnight scheduling job. Unsupervised classification is used to detect outlier without prior knowledge, which captures emerging behaviors of frauds which do not appear in historical behaviours.
Deliverables & Benefits
With our Advanced Fraud Detection (AFD) engine, B is now able to:
- more accurately detect frauds,
- enhance customer satisfaction,
- improve operation procedures, and
- develop a comprehensive and scalable risk management process.
- More efficient loss recovery
By looking at several key metrics, B is able to accurately identify fraudulent transactions in advance which enables the bank to prevent loss. The system was 80% accurate in identifying frauds, leading to a significant lift of customer satisfaction.
- Scalable Loss Prevention Plan
The AFD engine provides a powerful and flexible foundation on which B can expand its loss prevention efforts. With the detection engine, it can trim the alert volume according to its risk management capacity.
- Early Prevention of Risk
B uses the detection engine to explore the customer profile to uncover and analyze mule accounts, which are employed by the fraudsters to receive the illicit money, not only those victim accounts. Therefore, frauds can usually be prevented before being committed.
- Improved Operation Procedures
B has been able to use the AFD engine to explore the big data to uncover and visualise anomaly behavior patterns, which indicates interesting knowledge beyond the expertise of domain experts. For example, the investigation against suspicious transaction could be optimized in information collection according to the personalized behavior patterns.