Today deep learning is taking its rise in hydrometeorological applications, and it is critical to extensively evaluate its prediction performance and robustness. In our study, we use deep all convolutional neural networks for radar-based precipitation nowcasting, which has a crucial role for early warning of hazardous events at small spatiotemporal scales. Our trial and error study focuses the particular importance of selecting and adopting suitable data preprocessing routine, network structure, and loss function regarding input data features, and, as a result, highlights limited transferability of methods in existing studies. Results show that parsimonious deep learning models can forecast a complex nature of a short-term precipitation field evolution and compete for the state-of-the-art performance with well-established nowcasting models based on optical flow techniques.