US researchers developed a machine learning technique that predicts hit songs with 97% accuracy. With thousands of songs released daily, streaming services and radio stations face the challenge of selecting which songs will appeal to a wide range of audiences.