Wells Fargo plans to install a sophisticated computer system that uses artificial intelligence software to detect credit card fraud.
Wells, the 12th-largest card issuer, with $3.7 billion in credit card loans outstanding, according to a 1992 study by American Banker, plans to have the system operational for its entire card portfolio in the fourth quarter of 1993.
Wells declined to say whether fraud at the institution had risen significantly in recent years.
$2 Billion in Losses in '92
Industry observers say bank losses worldwide from credit card fraud in 1992 were at least $2 billion, and are expected to reach at least $2.4 billion in 1993.
The software Wells is installing is called a neural network, a type of artificial intelligence programming designed to discern patterns in large amounts of data in a way that mimics the human brain's nervous system.
The software, which is produced by HNC Inc., San Diego, can change the weights it assigns to variables automatically, as it learning from experience.
Up to $2 Million to Install
Few banks have installed these systems, which some bankers consider costly and unproven. Such systems cost between $250,000 and $2 million to install, according to HNC.
HNC says that Wells should be able to reduce fraud by 15% to 20% in the first year of operation.
A handful of banks have installed such systems, including Mellon Bank Corp., which has installed software from another vendor, Nestor Inc. in Providence, R.I.
Companies that have installed the HNC software, which is called Falcon, include: First USA Inc., Dallas; Colonial National Bank USA, Wilmington, Del.; and the issuer of the General Motors card, Household Credit Services, a unit of Household International, Prospect Heights, Ill.
|False Positives' a Problem
Critics of neural network technology say that it is difficult for operators of the system to understand how the system reaches a conclusion, making it difficult to evaluate and adjust the system's decisions.
Others say that neural networks are more accurate than other kinds of fraud detection systems, because they can discern patterns in data that are not linked in a logical fashion.
Bankers say the biggest problem with fraud detection systems generally is that they frequently identify fraud where none has occurred. Depending on how a bank handles the information, cardholders could be inconvenienced if a suspect card is deactivated.
Ken Jones, marketing manager of HNC, says that most bankers are comfortable with 50 misidentifications of fraud for every real fraud that is identified. But the software can be customized to make a greater or fewer number of "false positives," Mr. Jones said.
Mr. Jones said that one institution, which he declined to identify, was saving $3,000 a day in averted fraud using the software.
Wells, like most of the other banks, will update the software every night. The software can also be updated during the day in order to give credit officers the opportunity to issue stops on transactions during the course of the day.
The software that Wells is installing is programmed with two types of data: data on fraudulent transactions culled from the general population, and purchase patterns of individual accounts. The system can theoretically learn how a given cardholder normally behaves, and can pinpoint behavior that deviates from that norm.