Data on institutional ownership of securities is nowadays publicly available in a number of jurisdictions and can thus be used in models for the prediction of security returns. A number of recently launched investment products explicitly use such institutional ownership data in security selection. The purpose of the current study is to apply statistical learning algorithms to institutional ownership data from the United States, in order to evaluate the predictive validity of features based on such institutional ownership data with regard to future security returns. Our analysis identified that a support vector machine managed to classify securities, with regard to their four-quarter forward returns, into three bins with significantly higher accuracy than pure chance would predict. Even higher accuracy was achieved when predicting realized, i.e. past, fourquarter returns.