Complex models for regression and classiﬁcation have high accuracy but are unfortunately no longer interpretable by users. We study the performance of generalized additive models (GAMs), which combine single-feature models called shape functions through a linear function. Since the shape functions can be arbitrarily complex, GAMs are more accurate than simple linear models. But since they do not contain any interactions between features, they can be easily interpreted by users. We present the ﬁrst large-scale empirical comparison of existing methodsforlearningGAMs. Ourstudyincludesexistingsplineand tree-based methods for shape functions and penalized least squares, gradient boosting, and backﬁtting for learning GAMs. We also present a new method based on tree ensembles with an adaptive number of leaves that consistently outperforms previous work. We complement our experimental results with a bias-variance analysis that explains how different shape models inﬂuence the additive model. Our experiments show that shallow bagged trees with gradient boosting distinguish itself as the best method on low- to medium-dimensional datasets.