Real-time, large-scale data streams from Twitter/X offer valuable insights for political scientists analyzing publicly expressed stances during election campaigns. However, the fast-moving nature of social media events and topics on social media leads to concept shifts that can deteriorate model performance. To the best of our knowledge, we present the first dataset for multi-target stance detection toward German political actors, collected from X. The dataset is temporally fragmented and designed to support the investigation of continuous model adaptation and evaluation. The dataset comprises Twitter/X posts related to the 2022 German state election in North Rhine-Westphalia. The annotation procedure captures stances toward political parties and their leading candidates. To establish baselines, continuous retraining techniques for temporal model adaptation across different model sizes were investigated. Models of different scales were initially trained and then evaluated on future tweets. These models were subsequently retrained continuously on new data over the final eight weeks of the election campaign to assess the impact of different strategies across three model sizes. The results show that the benefits of continuous retraining diminish as model size increases. An interesting side result of this work is evidence that the language used in tweets becomes more emotional as election day approaches.