Neural network proves potent ally in war against credit card fraud.

Before the installation of a neural network system seven months ago, Mellon Bank Corp. officials were spinning their wheels when it came to combating credit card fraud.

The were forced to investigate up to 1,000 transactions each day that their computer system had flagged as possibly fraudulent. They would pare down the list and follow up with about 150 phone calls a day to cardholders to make sure the transactions were legitimate.

For their efforts they turned up only one fraudulent transaction a week.

Needle in a Haystack

"We weren't doing very well," said Barbara Donovan-Lamb, the fraud prevention manager for Mellon Bank's credit card unit in Wilmington, Del. "It was like trying to find a needle in a haystack."

But since Mellon's 1.6 million credit card accounts have been monitored by the system provided by Nestor Inc. of Providence, R.I., efficiency has improved markedly.

The more discerning system, using a type of artificial intelligence to make more accurate preductions, flags only about 400 accounts per day. Bank investigators reduced the number of follow-up calls to about 120. And they've significantly improved the rate at which they identify fraud, nailing four cases a day -- 28 times the previous rate.

Nestor's neural network has also helped Mellon to sharply cut the amount of time an account stays blocked between being flagged and being assigned to an investigator -- from 11 days to two.

Although she would not specify the amount Mellon loses annually to fraud, Ms. Donovan-Lamb said that the bank expects a 40% reduction for calendar 1993 as a result of the new system.

Similarly, Colonial National Bank, the credit card unit of Advanta Corp., Horsham, Pa., reports it has reduced losses due to nonreceipt of cards and counterfeiting by as much as 50%, since the installation of the Falcon neural detection system, offered by HNC Inc. of San Diego.

The still-developing technology of neural networks -- computer systems that work and "think" much as a human brains do -- has been revolutionizing the credit card industry by pro-viding fraud detection that is far more complex and effective than former processes.

Heavy Toll

The results reported by Mellon, Colonial, and others offer a bright ray of hope at a time when fraud losses have been casting a shadow in the industry. The Nilson Report of Oxnard, Calif., estimates fraud will cost the industry at least $967 million this year, excluding bankruptcies.

Neural technology has emerged from the research and development stage and has been used for various purposes in the financial industry. Like the organic neurons of the brain, the electronic components of a neural network can "learn" by example.

The credit card fraud investigators who use the systems feed hundreds or thousands of examples of both regular account activity and fraudulent situations into their computers, in effect "teaching" the systems to differentiate between regular activity fluctuations and possibly illegal transactions.

Enhancing Accuracy

"Neural nets just feed off of data," said Ken Jones, HNC's marketing communications manager. "And the more information you have, the more accurate your models are going to be."

Doug Reilly, vice president of financial services at Nestor, said the traditional statistical computer model is a simpler, more primitive version of a neural network. "You can regard a statistical model as a one-cell neural network, whereas a neural network you can train to have hundreds and thousands of cells," he said.

The more traditional systems flag any account activity that exceeds certain parameters. For example, Mellon Bank's former rule-based system would flag any account that had more than four transactions in one day. This did not account for sudden or brief alterations in spending patterns, due to vacation or holidays, and was not tailored to suit the varying activities of cardholders.

Neural networks track transactions, grading them according to the risk of fraud, and build a score on each account -- the higher the score, the greater the chance of fraud.

Ms. Donovan-Lamb said Mellon catches most of its fraud cases in the lower range of scores. She also pointed out that the bank was able on several occasions to alert customers that their cards were missing before they had realized it.

The neural networks themselves vary in terms of complexity -- how many passes, or examples, the system needs before it starts to "learn," how wide the range it scores the account within, and how many models it uses to base its decisions on.

HNC, the largest neural hardware and software producer, uses two models on which each account is rated: The first is a general model based on cumulative fraud data from all HNC'S clients; the second is a model based on the individual cardholder's account and his or her particular activities.

Credit Card Issuers

Some of the fastest-growing credit. card issuers already use Falcon, including Colonial National Bank, Household Credit Services, and First USA Bank, as well as First Data Resources, the largest third-party processor. HNC is also implementing systems for Wells Fargo Bank and AT&T Universal Card Services.

Falcon alone currently examines close to 25 million accounts for nine major clients. That number will almost triple after a majority of First Data Resources' 56 million accounts enter the fold in 1994.

HNC started in network research, working on government contracts, in 1986.

"Falcon is a case study of the successful conversion of military technology to commercial applications," according to Robert L. North, HNC'S president and CEO. "The product's neural networks, which were initially used to detect armored tanks hidden among boulders and trees. now detect fraudulent transactions hidden among vast amounts of valid transaction data."

The Falcon service costs between $250,000 and $2 million, depending on the size of the account portfolio, the company's fraud losses, and the length of the agreement. Systems take about 30 to 60 days to implement and typically pay for themselves in a year or less, according to Mr. Jones and fraud detection officials who use neural networks.

NeuralWare, a Pittsburgh-based neural software producer, offers courses in process control and financial applications for neural networks, in what its officials lovingly call a "geek school."

According to Jane Klimasauskas, NeuralWare's vice president of sales and marketing, the classes have been so well received that business leaders will wait on line for hours for an opening.

The most popular course, in financial forecasting, costs $5,000 for four days and closes out with 22 people. She estimated that about 40% of students are interested in using neural networks for business applications.

"What we're finding is an interest not only in neural nets, but also genetic algorithms and fuzzy logic ... and making hybrid systems." Ms. Klimasauskas said.

Trading Application

Indeed, the meteoric rise of neural networks in business has already extended beyond credit card fraud detection. One of NeuralWare's clients, an executive for the World Bank, uses the technology for trading.

Neural networks also promise to become a powerhouse technology in other financial areas, such as mortgage appraisal and debit card fraud.

HNC and Visa International have struck a deal to build a merchant fraud detection system. Gail Murayama, manager of public relations for Visa International, expects that this network could be instituted by late 1994.

This would not only add another dimension to Falcon's existing cardholder fraud detection and secure accounts at issuing banks, but inform acquiring banks on merchants' backgrounds and fraud records.

"There are merchants that are more prone to fraud," Ms. Murayama said. "And the networks can learn to raise a little red flag to say this may be questionable."

For reprint and licensing requests for this article, click here.
MORE FROM AMERICAN BANKER