This thesis aims to improve the current systems to identify fraudolent bank trans- actions. We start by analyzing the state of the art, in particular the work done in the BankSealer project, focusing on the temporal analysis of user profiles. We then suggest a new approach at identifying anomalous shifts in user spending patters, that should prove more general and effective than the current one.
After having defined a possible new system, we test its effectiveness and generality with a fully working prototype implementation against a set of over a million real bank transactions, and conclude by explaining the main success areas of the new approach, and the possible areas of improvement for future work.