Falcon: a neural network against card fraud
In September 1992 HNC of San Diego introduced Falcon, a system that scored the probability of fraud with a neural network while each credit card transaction was being authorised. According to the company's annual report, by August 1994 Falcon had been bought by 18 of the 20 largest US credit card issuers and monitored over 100 million accounts.
Why it matters
A neural network became part of everyday payments: it scored a transaction during authorisation, while the customer was still waiting for an answer. The company believed Falcon was the first commercial system to detect fraud in real time.
Falcon combined a general fraud model, trained on data from many banks, with a separate profile of each cardholder. American Banker of 26 October 1993, reporting a contract with AT&T Universal Card Services, wrote that Falcon was already employed by many of the nation's largest issuers. In 1994-1996 Falcon and related products made 36 to 60 percent of HNC's licence revenue; by early 1997 the company held contracts to monitor more than 150 million cards. What the record does not claim. The figures of 18 of 20 and 100 million are the company's own words in a filing to the regulator; it attributes the ranking of issuers to the trade bulletin The Nilson Report, which was not read. The independent newspaper gives no count of cards, so confidence is medium. The report's statement that Falcon saved its users over 100 million dollars in a year is not repeated here as fact.