Although extremely popular in the industry due to their high predictive performance, the lack of interpretability of machine learning algorithms raises concerns from practitioners and regulators. We propose an original model-agnostic method aiming to unravel the opacity regarding black boxes' decision process. Specifically, the method mesures the contribution of input features to the predictive performance of a model through the decomposition of an evaluation metric. The approach can be applied to any type of model, econometrics or machine learning, and to a wide class of evaluation metric including the most famous measures such as the R2 for regression problem or the Area Under the ROC Curve (AUC) for classification problems. We show the properties of our decomposition method and illustrate how feature contributions are estimated, even in a high-dimensional model context. A framework for local analysis is also developed. We highlight the usefulness of the approach using real data of credit scoring applications. Through the decomposition of several evaluation metrics, we illustrate how the decomposition can be used in practice to improve the decision making associated to black boxes.