Stance detection identifies a text’s position or attitude toward a given subject. A major challenge in Roman Urdu is the lack of a publicly available dataset for political stance detection. To address this gap, we constructed a high-quality dataset of 8,374 political tweets and comments using the Twitter API, annotated with stance labels: agree, disagree, and unrelated. The dataset captures diverse political viewpoints and user interactions. For feature representation, we employed TF-IDF due to its effectiveness in handling high-dimensional, context-sensitive Roman Urdu text. Several machine learning classifiers were evaluated, with Random Forest achieving the highest accuracy of 95%. Additionally, we fine-tuned the transformer-based RoBERTa model, which outperformed traditional methods with 97% accuracy. Our results demonstrate the potential of combining machine learning and deep learning for stance detection in low-resource languages. This study not only introduces a novel dataset but also provides a robust evaluation of methods, highlighting the importance of modern AI techniques in processing informal and multilingual text data.