Abstract: The popularity of sentiment and emotion analysis has lead to an explosion of datasets, approaches, and papers. However, these are often tested in optimal settings, where plentiful training and development data are available, and compared mainly with recent state-of-the-art models that have been similarly evaluated. In this paper, we instead present a systematic comparison of sentiment and emotion classification methods, ranging from rule- and dictionary-based methods to recently proposed few-shot and prompting methods with large language models. We test these methods in-domain, out-of-domain, and in cross-lingual settings and find that in low-resource settings, rule- and dictionary-based methods perform as well or better than few-shot and prompting methods, especially for emotion classification. Zero-shot cross-lingual approaches, however, still outperform in-language dictionary induction.