This paper presents the results of a real-world evaluation comparing different machine learning approaches for water consumption forecasting. The data sets stem from a public water supplier of the city of Klagenfurt in Austria. Apart from prediction accuracy, which is clearly the most fundamental criterion for a forecasting model, we also investigate aspects of model interpretability, as well as seasonal and geographic model robustness, that are of high practical importance for public water suppliers. To this end the selected forecasting methods differ in their level of complexity and interpretability. Furthermore, data sets that differ seasonally and geographically allow to draw conclusions about method robustness. It turns out that also rather simple methods provide a high degree of accuracy whilst also being interpretable, robust and computationally efficient.