The prediction of compound properties from chemical structure can be improved by correcting for an expected structure-specific calculation error, which is derived from the method performance with structurally similar compounds selected from a data set with validated experimental values. For the automatic identification of proper reference compounds by a computer program, a suitable algorithm is required. Two different approaches have been tested for various group contribution methods to calculate water solubility and octanol/water partition coefficient of organic chemicals. One algorithm is based on a hierarchical classification scheme, and the other evaluates structural similarity through comparison of atom-centered fragments. The latter algorithm has proven to be superior.