A novel filter feature selection technique is introduced. The method exploits the information conveyed by the evolution of the training samples weights similarly to the Adaboost algorithm. Features are selected on the basis of their individual merit using a simple error function. The weights dynamics and its effect on the error function are utilised to identify and remove redundant and irrelevant features. In experiments we show that the performance of commonly employed learning algorithms using features selected by the proposed method is the same or better than that obtained with features selected by the traditional state-of-theart techniques.