Sequential fraud detection by determining proper sequence length in payment cards using HMM

The use of bank cards has increased significantly in recent years. This has resulted in increasing the probability of internet payment card frauds and has highly imposed losses on customers, institutions and banks. The methods used to detect frauds in this area mainly require a huge volume of historical data. On the other hand, these methods usually work well when there are single bank transactions, which means they only have the ability to detect frauds during single bank transactions and do not reveal fraudulent sequence identification.In this paper, a model is proposed to determine the appropriate sequence length required to evaluate every single customer's spending behavior. Through adding the feature of fraudulent sequence detection in payment cards, the proposed model has been completed. This model automatically creates and updates the Hidden Markov Model of each sequence, and ultimately detects frauds by comparing the Kullback-Leibler divergence between Hidden Markov Model of each sequence. The fraud detection is presented by real semi-supervised payment cards data of an Iranian bank. The obtained F-Score, derived from 7 real fraudulent scenarios created under the supervision of a bank expert, representing 87%. Using the proposed model also leads to a reduction in the fraudulent sequences incidence cost of 81%.