One of the principal inconveniences that analysis and information processing presents is that of the representation of dataset. Normally, one encounters a high number of samples, each one with thousands of variables, and in many cases with irrelevant information and noise. Therefore, in order to represent findings in a clearer way, it is necessary to reduce the amount of variables. In this paper, a novel variable selection technique for multivariable data analysis, inspired on stochastic methods and designed to work with support vector machines (SVM), is described. The approach is demonstrated in a food application involving the detection of adulteration of olive oil (more expensive) with hazelnut oil (cheaper). Fingerprinting by H NMR spectroscopy was used to analyze the different samples. Results show that it is possible to reduce the number of variables without affecting classification results.