Developing models to prevent financial fraud is a complex task. To help suppress financial fraud we used innovative techniques such as anomaly detection. Based on the Markov Chain approximation, we developed a model to analyse sequential patterns on a given data set. Stand-alone features built from this approximation were one part of the complex model implemented and used in production.
In recent years, enormous strides in reducing fraud rates were made thanks to Artificial Intelligence. Nonetheless, fraud is not completely suppressed and developing reliable anti-fraud models requires going beyond traditional methods. We adapted the Markov Chain approximation to spot client behaviour anomalies using sequential web-collected data. The study is based on data taken on a time span of six years. Behavioural patterns were investigated and future sequential patterns were predicted based on the frequency of similar sequential events in the past.
Several features based on these probabilities were used as an important part of the complex model in production, with the goal of isolating fraudulent client behaviour based the sequential information. Alternatively, this approach can also be used to predict the immediate next event on a webpage.
In this talk, we are going to present the general approach, followed by a comprehensive walk through the modelling steps, as well as details on tools and implementation. Other possible applications areas are going be discussed as well.