We propose a holistic system to classify ancient Roman Republican coins based on their reverse-side motifs. The bag-of-visual-words (BoW) model is enriched with spatial information to increase the discriminative power of the coin image representation. This is achieved by combining a spatial pooling scheme with co-occurrence encoding of visual words. We specifically address the required geometric invariance properties of image-based ancient coin classification, as coins from different collections can be located at differing image locations, have various scales in the images and can undergo various in-plane rotations. We evaluate our method on a data set of 2,224 coin images from three different sources. The experimental results show that our proposed image representation is more discriminative than the traditional bag-of-visual-words model while still being invariant to the mentioned geometric transformations. For 29 motifs, the system achieves a classification rate of 82%. It is considered to act as a helpful tool for numismatists in the near future, which facilitates and supports the traditional coin classification process by a faster presorting of coins.