One of the fundamental problems in the development of prosthetic fingers is the recognition of finger movements using surface electrocardiogram (EMG) data. Most of the previous studies have proposed the classification of EMG signals using features curated using expert knowledge. We here consider automatic generation and selection of EMG signal features without relying on domain knowledge. We then develop a classification method based on random forests. Our results show that the proposed method achieves 97.5% accuracy. We also present a discussion of the features which are important in distinguishing finger movements.