Neurofeedback therapies are an emerging technique used to treat neuropsychological disorders and to enhance cognitive performance. The feedback stimuli presented during the therapy are a key factor, serving as guidance throughout the entire learning process of the brain rhythms. Online decoding of these stimuli could be of great value to measure the compliance and adherence of the subject to the training. This paper describes the modeling and classification of performance feedback potentials with a Brain-Computer Interface (BCI), under a real neurofeedback training with five subjects. LDA and SVM classification techniques are compared and are both able to provide an average performance of approximately 80%.