Effective and cheap methods for detecting fraud and, guaranteeing wine authenticity, are of paramount importance in the sector. In this sense, three different kinds of prediction models (random forest, artificial neural networks, and support vector machines) were developed to classify wines, according to their element contents (metals and metalloids, obtained using an inductively coupled plasma with a quadrupole mass spectrometer, and an optic emission spectrophotometer). One models were developed using 45 inputs variables, and then the models were subjected to a process of reducing variables to simplify models and save material and time costs. A total accuracy was reached in all phases for the white wines-random forest models. From a practical point of view, the accuracy and the errors obtained by the selected models (except for red wines-artificial neural network developed using reduced variables) are acceptable. The models developed with fewer variables, can make the prediction task easier.