Stance detection plays an important role in contemporary news analysis by identifying the argumentative relationship between a claim and its associated textual context. In the era of algorithm-driven media, news articles often convey implicit support, opposition, or neutral discussion of specific claims, making stance analysis essential for detecting media bias and researching misinformation. However, accurately modeling such relations remains challenging due to long document lengths, implicit stance expressions, and complex discourse structures. This study evaluates an end-to-end Transformer-based stance detection approach that fine-tunes the DeBERTa-base language model on news text using the Fake News Challenge Stage 1 (FNC-1) dataset under a stance-relevant formulation. The proposed framework updates all parameters of the pre-trained model directly during training, avoiding handcrafted feature engineering and auxiliary classifiers. Claim–context pairs are jointly encoded and formulated as a three-class stance classification task (agree, disagree, discuss), following the exclusion of unrelated instances to focus on argumentative relations. To ensure robust evaluation under class imbalance, model performance is assessed on a held-out test set using standard classification metrics. Experimental results on the test data show that the proposed approach achieves 96.28% accuracy and 96.23% F1-score, indicating balanced precision–recall performance across stance categories. These findings suggest that a carefully configured end-to-end fine-tuning strategy based on DeBERTa-base is effective for capturing argumentative relations in news text within a three-class stance-relevant setting, providing a reliable and reproducible solution for document-level stance detection without relying on complex architectural modifications or feature engineering.