Classification of electroencephalograph (EEG) signals is the common theoretical background of various recognition tasks, such as EEG-based diagnosis of diseases, identification of sleep stages and the recognition tasks related to EEG-controlled spelling devices or web browsers. Projection-based classification of EEG is one of the state-of-the-art techniques to solve such tasks. In this paper, we propose (i) to utilize asymmetric loss linear regression for projection-based classification and (ii) to use genetic algorithm to select reference signals. We performed experiments on a publicly available EEG dataset. Our model aimed to classify patients according to a disease (alcoholism). The results show that both proposed techniques may increase accuracy.