The users of information systems often have to deal with outliers in their data. Such outliers can have negative (i.e. abnormal observations) or positive (i.e. detection of new features) impact on their work. Despite the fact, that several methods of outlier detection already exist, there is still a need to improve them. In this work we propose a method for evolutionary outlier detection. The novelty of our approach is a set of criteria, which are used to decide, whether to treat observation as an outlier or not. Conducted research revealed that our method performs very well on the selected problems.