“simple models can close the accuracy gap when applied to large data sets”