8.3 Credit-Card Fraud

Years before the Web, credit-card fraud was already a big problem, where even the simplest instance of fraud was difficult to detect in light of the millions of transactions taking place every minute of every day. Credit-card issuers had been combating fraud for several years both by hiring teams of data miners and developing in-house detection systems employing pattern-recognition technologies, such as neural networks, or by outsourcing the problem to outside experts.

Prior data mining analyses have found that certain variables contain telltale clues for detecting credit-card fraud. Citibank, for example, found that it could dramatically reduce fraud when it passed certain variables though a neural network. This included the following four information items:

The Standard Industry Code of the product or service being purchased
The number of transactions for that day
The dollar amount of the transaction
The zip code of the transaction

An amusing trend that Citibank discovered was that repeated purchases of expensive Italian shoes in New Jersey tended to indicate a high probability that they were being made with stolen credit cards. This example leads to some important lessons, which can be applied to the problem of on-line fraud. Features such as location, the type of product or service being purchased, the velocity of the transactions, and the dollar ranges are all key indicator for detecting and deterring on-line fraud. Another red-flag indicator for on-line sales is a transaction where the shipping address is different from the billing address. The experience of several e-retailers indicates that this type of transaction should be carefully monitored.