First USA Bank, one of the nation's top credit card banks, has become the first U.S. financial institution to install a neural network to combat credit card fraud.
Neural networks - programs that imitate the workings of the human nervous system - can detect complex patterns in large amounts of data.
Several credit card banks that have tested the technology, including Mellon Bank Corp., Pittsburgh, and First USA, Wilmington, Del., have said the technology is useful in pinpointing when a credit card transaction submitted for authorization varies from from the cardholder's established purchasing pattern.
Since these deviations are one of the most reliable indicators of a fraud, neural networks afford banks the opportunity to give extra scrutiny to suspicious transactions before issuing an authorization, experts said.
While critics maintain neural networks are expensive and unproven, the few banks that have looked at using the technology believe it could prove a useful tool.
For 1991, MasteCard and Visa reported that fraud losses in the United States reached $500 million, a 40% increase over the 1990 level. More than half a million cases of fraud were detected, in 1991, according to the bank card associations.
"A neural network will identify a fraudulent transaction sooner and better than any manual judgmental process ever could," said Edward Scully, vice chairman of First USA
First USA completed testing the system from HNC Inc., San Diego, in September.
So far, the system is mainly a productivity tool. The number of accurate fraud identifications is about the same as under the old system of human analysis. But the bank has reduced the number of accounts reviewed by about 30% since moving all their transactions through the neural network, Mr. Scully said.
The bank is currently processing transactions in batches, which enables them to investigate potential fraud only after a transaction has been completed.
While this method is useful for stopping repeated fraud with the same credit card, Mr. Scully said that the bank hopes to be moving beyond batch processing by the spring so that operators at the point of sale can be alerted to potential fraud before a transaction is finished.
The neural network system at First USA cost "in the low six figures," to install, Mr. Scully said.
While use of the technology in banking is in its infancy, First USA is not alone in its belief that neural networks can be an effective weapon in the war on credit card fraud.
Mellon Bank, is in the late stages of testing neural network software from Nestor Inc. of Providence, R.I., and expects to begin running live work by the middle of December. Mellon predicts the system will reduce fraud losses by 15 to 20%.
Spotting Stolen Cards
Experts said neural networks - which involve integration and customization costs of about $200,000 for a bank with more than $2 billion in its credit card portfolio - are effective in identifying a number of types of fraud transactions, including those resulting from new cards stolen while in transit.
The 1992 ABA report on the bank card industry found that nearly 20% of all fraud dollar losses are related to such thefts.
However, experts said neural networks would be most useful in preventing losses on cards lost or stolen from customers. which accounts for more than half of all card fraud in the United States, according to the ABA.