The detection of bank frauds is a topic which many financial sector companies have invested time and resources into. However, finding patterns in the methodologies used to commit fraud in banks is a job that primarily involves intimate knowledge of customer behavior, with the idea of isolating those transactions which do not correspond to what the client usually does. Thus, the solutions proposed in literature tend to focus on identifying outliersor groups, but fail to analyse each client or forecast fraud. This paper evaluates the implementation of a generalized linear model to detect fraud. With this model, unlike conventional methods, we consider the heterogeneity of customers. We not only generate a global model, but also a model for each customer which describes the behavior of each one according to their transactional history and previously detected fraudulent transactions. In particular, a mixed logistic model is used to estimate the probability that a transactionis fraudulent, using information that has been taken by the banking systems in different moments of time.