The goal of automated machine learning is to search for a suitable machine learning model given a problem specified by data. In this work we are interested in the domain of complex supervised machine learning models. Since the performance evaluation of complex models is time consuming, several performance prediction approaches have been proposed. The objective of this work is utilizing the meta-learning approach – we make use of the previous experience with models’ performance on benchmark tasks in order to estimate the performance on a previously unseen dataset. We extract features describing the structure of the models, as well as meta-data features describing the datasets. The information about the model and task are aggregated and used to train a regression model serving as a performance predictor. The advantage of this approach is that the model performance is estimated without training, using just feature extraction and regressor inference. The approach is successfully tested on a large collection of models and their performance information from the OpenML repository, and the results are reported in the paper. Several alternatives for feature extraction and regression procedures are also compared.