8.5 The Risk Scores

Because of the broad profiles, however, a transaction is not so easily categorized into one of the two groups. Therefore, fraud-detection systems must also assign a risk score to each transaction. This score represents just how much the system felt the transaction was fraudulent. This is a continuous value number. Fraud profiles can be developed and evolved, providing a basis for fraud detection using artificial AI. Neural networks and machine learning are commonly used today to evaluate each transaction and assign this type of risk score.

This risk score is compared against a predetermined score threshold, thereby enabling acceptance or the firing of an alert rejection in real time. For example, a risk score of 90 might signal a fraudulent purchase, enabling any transaction with a score of 89 or below to be authorized. However, risk scores are also based upon previous risk scores for each individual. This means that a score of 89 will pass the test, but a subsequent score of 89 will be bumped up due to the previous score and the purchase will not be authorized. Also, the personal risk score of individuals who on average run up thousands of dollars on their cards each month may be significantly higher than those who average only a few dollars.

Seasonality can also hinder the detection of fraudulent purchases; credit-card fraud activity can increase 15% or more during the holiday season. High rates of purchase can put an extra load on a fraud-detection network system. Unusual purchases for a profile, such as those of gifts for relatives, can also affect the accuracy of the system. In effect, a consumer's purchasing profile is thrown out the window during the holiday season. Shopping patterns are unpredictable during this time, and thus, credit-card companies tend to treat this loss as a cost of doing business.