Intelligence systems for fraud detection

Increasingly, “intelligent” software techniques such as neural networks, genetic algorithms, and fuzzy logic are incorporated into automated systems designed to detect and highlight instances of fraud. These techniques have been applied to identify credit card fraud, insurance claims fraud and insider dealing and are also used in company auditing. In this article Jason Kingdon of Searchspace Limited in London introduces intelligent systems and examines some of their fraud-prevention applications.

A version of this article originally appeared in “Genetic Algorithms and Fuzzy Logic Systems” edited by Elie Sanchez, Takanori Shibata and Lotfi Zadeh. World Scientific Publishing Co Ltd, 1997.

Detecting fraud

There are many reasons why most service sector companies do not have comprehensive fraud detection systems. The technical difficulties are significant, centring on the vast number of records that need to be intelligently examined, cross-checked, and verified. In most instances the sheer volume of transactions presents the real difficulty in developing an automated processing system, the essence of which must be speed as well as accuracy. Comprehensive checks on each transaction impose a severe burden on the speed and efficiency of the system, equally too few checks are costly in terms of undetected fraud. In order to appreciate the scale of the “data problem” consider the following: Visa International has 16 million transactions a day for which information relating to the place of purchase, merchant type, transaction amount and time and date is recorded. This level of information through-put is typical for many service-based industries from telecommunications to the financial markets. For example, the London Stock Exchange processes an average of 35,000 bargains a day, in which each bargain has over 100 data fields attached. Any investigation into market abuse takes place within an information background consisting of millions of bargains, over 2500 different stocks, hundreds of brokers, and millions of clients. These very large volumes are the greatest barrier to an effective fraud screening process and in most cases completely exclude any real-time human based approach.