In this paper, we build on previous work on predictability in morphology, where entropy-based techniques have first been introduced to measure predictability in inflectional predictions, and then extended to predictability of the citation form of derivatives, and we apply those techniques to explore variation in the predictability of derivatives from different inflected forms of their bases. As a case study, we choose a paradigmatic system that includes the principal parts of Latin verbs and related agent and action nouns, where interesting variation can be found according to the forms considered for the predictions. Our results confirm many of the expectations that can be formulated on the basis of a qualitative analysis and additionally highlight less evident characteristics of the distribution of morphological patterns in our data. We argue that a unified approach like the one adopted in this paper is capable of capturing the complex and bidirectional interaction between inflectional and derivational information, and we point to interesting issues that arise from such a unified approach, highlighting their broader implications for morphological theory and computational modelling.