The thesis deals with the theme of machine learning applied to credit risk. Starting from a theoretical introduction of the tools that will be used in the course of the work, we get to observe the performance of 5 different algorithms for binary classification : Logistic regression, Naive Bayes, K-Nearest neighbors, Support Vector Machines and Ensemble trees. In particular, two different datasets are used: a synthetic one, and a real one. Then Ensemble Classifiers, which can be intuitively interpreted as a way to combine predictions generated by different algorithms, are introduced. Several ways of mixing prediction are observed, starting from hard- voting, which is simply a majority voting system, and coming up to more sophisticated methods like stacking and blending. Furthermore, an appreciable improvement in several measures of the classification performance is obtained.