Accumulating Evidence of Fraud Using a Continuous TimeHidden

This article presents a hidden Markov model designed to screen a telephone, computer , or other type of account for criminal activity. An augmented variables sampler is used to draw model parameters from their posterior distribution. Using estimated parameters , the forward-backward algorithm for hidden Markov models rapidly computes the probability of a criminal presence on the account at any point in time. The model allows temporally close transactions to share information about criminal activity. The model accepts covariate information produced by fraud detection procedures focusing on individual transaction characteristics, resulting in a principled accumulation of evidence over time. The model may be used to estimate the nancial loss the account has experienced due to fraud. The model is illustrated on a data set containing information about fraud in international telephone traac.