Optimizing a Fraud Detection Process

Nuno Homem, Joao Paulo Carvalho

Fraud in telecommunications services or financial transactions is a major problem as it impacts from 1% to 3% of the revenues. This is in most cases a customer specific behavior that companies need to detect in order to minimize it. Detecting specific types of behavior as soon as it happens is critical, and to that purpose companies deploy sophisticated detection systems. The biggest challenge to fraud detection systems is accurately predict in near real time that a customer is a fraudster or that is service is being used in fraudulent way. As this may happen to any customer at any time it is mandatory to monitor the entire customer base – sometimes several million customers making several transactions per day. Optimizing the detection process is therefore critical. To model and analyze the problem Finite State Automata and Markov Chains were used. To solve the optimization problem Dynamic Programming and Stochastic Hill Climber algorithms were chosen.

PDF full paper